<?xml version="1.0" encoding="utf-8" ?><rss version="2.0" xmlns:tt="http://teletype.in/" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:media="http://search.yahoo.com/mrss/"><channel><title>Khasky</title><generator>teletype.in</generator><description><![CDATA[Developer with founder delusions]]></description><image><url>https://img2.teletype.in/files/55/b4/55b4776f-bf63-4267-ac50-9a2d8049e612.png</url><title>Khasky</title><link>https://teletype.in/@khaskydev</link></image><link>https://teletype.in/@khaskydev?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><atom:link rel="self" type="application/rss+xml" href="https://teletype.in/rss/khaskydev?offset=0"></atom:link><atom:link rel="next" type="application/rss+xml" href="https://teletype.in/rss/khaskydev?offset=10"></atom:link><atom:link rel="search" type="application/opensearchdescription+xml" title="Teletype" href="https://teletype.in/opensearch.xml"></atom:link><pubDate>Mon, 16 Nov 2026 10:50:18 GMT</pubDate><lastBuildDate>Mon, 16 Nov 2026 10:50:18 GMT</lastBuildDate><item><guid isPermaLink="true">https://teletype.in/@khaskydev/practical-ai-coding-workflows-awesome-agent-skills</guid><link>https://teletype.in/@khaskydev/practical-ai-coding-workflows-awesome-agent-skills?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/practical-ai-coding-workflows-awesome-agent-skills?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Practical AI coding workflows with Awesome Agent Skills</title><pubDate>Sun, 18 Oct 2026 20:47:56 GMT</pubDate><media:content medium="image" url="https://img2.teletype.in/files/5c/1e/5c1e4844-3788-4418-88aa-2ebffbe47d5b.png"></media:content><category>AI</category><description><![CDATA[<img src="https://img4.teletype.in/files/38/c5/38c51cb5-84b8-41e1-b8c4-3db10021cd16.png"></img>An AI agent can finish a feature, pass the tests, and still leave behind unnecessary abstractions, misleading comments, or a README that no longer matches the product. Asking it to &quot;check everything&quot; gives it little guidance about what to investigate or what would count as proof.]]></description><content:encoded><![CDATA[
  <figure id="YzOs" class="m_original">
    <img src="https://img4.teletype.in/files/38/c5/38c51cb5-84b8-41e1-b8c4-3db10021cd16.png" width="1600" />
  </figure>
  <p id="BSAn">An AI agent can finish a feature, pass the tests, and still leave behind unnecessary abstractions, misleading comments, or a README that no longer matches the product. Asking it to &quot;check everything&quot; gives it little guidance about what to investigate or what would count as proof.</p>
  <p id="H5ot"><a href="https://github.com/khasky/awesome-agent-skills" target="_blank">Awesome Agent Skills</a> turns recurring engineering tasks into reusable procedures. Its 49 skills cover code review, debugging, testing, architecture, security, documentation, Git, and writing. Each defines a process and an expected result: a verified fix, a prioritized audit, a simpler plan, or documentation grounded in the code. I distilled each one from public collections and documented practice, then sharpened it in rounds of use on real repositories.</p>
  <p id="ltJa">Its strongest use case is maintaining an existing product: understanding what has accumulated, deciding what can be removed, and checking that implementation, tests, and public promises still agree. 🔍</p>
  <h2 id="8LTi">What Makes the Collection Useful?</h2>
  <p id="0Nzo">Each skill lives in a folder containing a <code>SKILL.md</code> file and, where needed, supporting references. The procedures are written in Markdown and use the agent&#x27;s existing tools. They ship no scripts. The repository describes installation for Claude Code, Codex, Gemini CLI, Cursor, and other agents supporting the Agent Skills format. More demanding procedures still depend on capabilities such as subagents or browser automation.</p>
  <p id="R5dt">Skills complement <code>AGENTS.md</code> and <code>CLAUDE.md</code>. Those files hold standing project rules. A skill gives the agent a method for a particular job.</p>
  <p id="dKNF">Three choices shape much of the collection:</p>
  <ul id="BEza">
    <li id="TKFD"><strong>Evidence-backed findings.</strong> A problem should be tied to code, a traced execution path, or a command result. Confirmed defects and assumptions are distinguished.</li>
    <li id="tugg"><strong>Separate investigation and remediation.</strong> Many audits produce a report before any source changes, letting the owner choose what to fix.</li>
    <li id="4SS6"><strong>Explicit limits.</strong> Reports identify what could not be checked. A release verdict is meaningful only within the assessed scope.</li>
  </ul>
  <p id="sYvT">These requirements make the output easier to evaluate. &quot;The architecture could be cleaner&quot; becomes a proposal naming the problematic dependency, the consequence, and the change that would address it.</p>
  <h2 id="IHXN">How It Compares with Popular Engineering Skill Sets</h2>
  <p id="Vc1x">A useful comparison needs both popularity and overlapping work. The alternatives below have substantial GitHub followings and contain engineering skills or workflows for coding, testing, debugging, review, or auditing. General document collections, single behavioral prompts, and directories of links are outside this comparison.</p>
  <pre id="KLn5">Collection                         Stars   Relevant strengths
Superpowers                        301k    planning, TDD, systematic debugging,
                                           isolated implementation, review
Matt Pocock&#x27;s Skills               295k    requirements interviews, TDD, code review,
                                           architecture improvement, issue-driven work
ECC / Everything Claude Code       283k    engineering skills and specialist agents,
                                           verification, security, hooks, memory
Addy Osmani&#x27;s Agent Skills         108k    full development lifecycle, behavior-preserving
                                           simplification, quality review, security, performance
Agentic Awesome Skills              48k    broad installable library with discovery,
                                           selection and stack-validation tooling
wshobson/agents                     41k    modular plugins for review, debugging, testing,
                                           refactoring, security, specialist development
Trail of Bits Skills               7.6k    security audits, differential review, static
                                           analysis, mutation testing, spec compliance</pre>
  <p id="l1hc">Star counts are approximate, as of 2026-10-17. Stars establish public visibility, not engineering quality. Trail of Bits is included as a particularly close audit specialist, rather than as one of the largest general-purpose libraries. Agentic Awesome Skills is broader than the others but supplies actual installable skills, making it relevant when choosing between a focused set and a large catalog.</p>
  <h3 id="1DTf">Compared with Matt Pocock and Addy Osmani</h3>
  <p id="BOaN">These are close alternatives, not merely popular repositories in the same ecosystem. Matt Pocock includes architecture surveys, disciplined debugging, TDD, and code review. Addy Osmani includes a dedicated simplification procedure that preserves behavior, respects project conventions, and verifies changes incrementally.</p>
  <p id="TVk8">Awesome Agent Skills is worth considering when you want the investigation divided into specific deliverables: a ranked test-gap worklist, a plan review counting what can be removed, a cleanup report, or a public-claims audit. Matt&#x27;s collection is particularly relevant when requirements clarification and domain modeling are your recurring problems. Addy&#x27;s fits when you want quality practices organized across the development lifecycle.</p>
  <p id="ao25">The distinction is the shape and coverage of the procedures. Simplicity, testing, and evidence are shared priorities.</p>
  <h3 id="SdS2">Compared with Superpowers, ECC, and wshobson</h3>
  <p id="youv">Superpowers provides a coordinated process from design to implementation and review. ECC combines engineering guidance with runtime integrations. wshobson packages specialist agents, skills, and commands into plugins, including developer essentials, code refactoring, security scanning, and multi-reviewer workflows.</p>
  <p id="AnLo">Awesome Agent Skills fits a team that wants to keep its existing process and add narrowly chosen investigations. Its review jury also has a specific design: isolated reviewers receive the same brief, then a separate judge cross-examines findings. By comparison, wshobson&#x27;s <code>multi-reviewer-patterns</code> allocates different review dimensions and consolidates the results. Both approaches serve multi-reviewer work. They organize independence differently.</p>
  <p id="BJio">I would choose the coordinated workflow or specialist plugin system when orchestration is the main requirement. I would choose individual audit procedures when a particular report is what you need.</p>
  <h3 id="DExK">Compared with Trail of Bits and Larger Libraries</h3>
  <p id="HHKZ">Trail of Bits is a strong alternative for security-focused work. It provides static-analysis workflows, vulnerability-pattern searches, false-positive verification, and dedicated testing procedures. Its specification-compliance skill also compares documentation with implementation, so that capability should not be presented as exclusive to Awesome Agent Skills.</p>
  <p id="zb84">The latter&#x27;s claims audit focuses on public product surfaces such as store listings, permissions explanations, language counts, and README commands, including repeatable checks and a separate correction phase. That is a concrete reason to consider it when maintaining both the product and its public descriptions. For deep vulnerability research and scanner-backed analysis, Trail of Bits is the more specialized collection.</p>
  <p id="8Lwj">Agentic Awesome Skills offers much broader discovery. A focused collection can reduce the selection work when your repeated tasks are already known. A large library is useful when you need to find capabilities across many stacks and domains.</p>
  <p id="PCdP"><strong>The practical reason to choose Awesome Agent Skills is its combination of maintenance procedures and explicit reports - not a demonstrated lead in coding performance.</strong> No verified head-to-head result here establishes that it fixes more bugs or works faster. Evaluate it on a familiar task. Did it find a real problem, supply sufficient evidence, avoid unwanted edits, and justify the time and context it consumed?</p>
  <h2 id="Gf4k">How to Use It Without Overcomplicating Your Workflow</h2>
  <p id="qucz">Start with the procedures you regularly need. For a bug fix, use debugging and test writing, then review the diff. For a cleanup, audit first, choose the changes, and apply them with behavior-preserving checks. For a new design, write the plan and review it for unnecessary mechanisms before implementation.</p>
  <p id="pCH3">An explicit request helps the agent select the intended procedure:</p>
  <pre id="xjfh">Use awesome-plan-simplify on this migration plan.
Check its assumptions against the code, identify what can be removed,
and explain the trade-off for each proposal.</pre>
  <p id="PSlw">The collection also includes writing and content skills: translation, grammar suggestions, document editing, adapting articles into posts, and publication workflows. Those are useful additions if you maintain both a product and its public material.</p>
  <h2 id="hWEe">Installation</h2>
  <p id="I66P">To install the collection globally for Claude Code, Codex, and Gemini CLI:</p>
  <pre id="oLek">git clone https://github.com/khasky/awesome-agent-skills.git
cd awesome-agent-skills
npx skills add ./skills -g -s &quot;*&quot; -a claude-code codex gemini-cli -y</pre>
  <p id="kRMT">You can select fewer skills or different agents. Claude Code also supports installation through the repository&#x27;s plugin marketplace:</p>
  <pre id="DejL">/plugin marketplace add khasky/awesome-agent-skills
/plugin install awesome-agent-skills@awesome-agent-skills</pre>
  <p id="iZQG">The README lists the same plugin route for Codex, GitHub Copilot CLI, Gemini CLI, Qwen Code, and Cursor. Choose one installation method to avoid duplicates. After installation, restart or reload the agent and confirm the skills are available. ✅</p>
  <p id="67J1">For updates, check whether the installed folders are copies or links. <code>git pull</code> reaches installed skills only when their links point to the clone. Copied installations need their own update step. The <a href="https://github.com/khasky/awesome-agent-skills#quick-start" target="_blank">repository README</a> explains the options.</p>
  <p id="jpsL">Awesome Agent Skills is worth trying when you can name a recurring maintenance task that needs a more precise method. Pick one, run its procedure on a familiar part of your project, and judge the report by its evidence, usefulness, and cost.</p>
  <h2 id="qBYh">References</h2>
  <ul id="DBNR">
    <li id="9qOd">Awesome Agent Skills repository: <a href="https://github.com/khasky/awesome-agent-skills" target="_blank">https://github.com/khasky/awesome-agent-skills</a></li>
    <li id="Ohou">Quick start and plugin install table: <a href="https://github.com/khasky/awesome-agent-skills#quick-start" target="_blank">https://github.com/khasky/awesome-agent-skills#quick-start</a></li>
    <li id="r2MH">Superpowers: <a href="https://github.com/obra/superpowers" target="_blank">https://github.com/obra/superpowers</a></li>
    <li id="GJhm">Matt Pocock&#x27;s Skills: <a href="https://github.com/mattpocock/skills" target="_blank">https://github.com/mattpocock/skills</a></li>
    <li id="KPlD">ECC / Everything Claude Code: <a href="https://github.com/affaan-m/ECC" target="_blank">https://github.com/affaan-m/ECC</a></li>
    <li id="Ksus">Addy Osmani&#x27;s Agent Skills: <a href="https://github.com/addyosmani/agent-skills" target="_blank">https://github.com/addyosmani/agent-skills</a></li>
    <li id="2kmD">Agentic Awesome Skills: <a href="https://github.com/sickn33/agentic-awesome-skills" target="_blank">https://github.com/sickn33/agentic-awesome-skills</a></li>
    <li id="2Glu">wshobson/agents: <a href="https://github.com/wshobson/agents" target="_blank">https://github.com/wshobson/agents</a></li>
    <li id="DIxc">Trail of Bits Skills: <a href="https://github.com/trailofbits/skills" target="_blank">https://github.com/trailofbits/skills</a></li>
  </ul>
  <hr />
  <p id="bIjq">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/non-english-prompts-cost-more-tokens</guid><link>https://teletype.in/@khaskydev/non-english-prompts-cost-more-tokens?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/non-english-prompts-cost-more-tokens?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Non-English prompts cost more tokens, context and money in most LLMs</title><pubDate>Mon, 05 Oct 2026 05:07:36 GMT</pubDate><media:content medium="image" url="https://img4.teletype.in/files/f5/8e/f58e42e1-2d58-4b8d-b853-c30ed0a34ee8.png"></media:content><category>llm</category><description><![CDATA[<img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/ftmahlzyyccpqoqglblg.png"></img>Claude Opus 5 costs $5 per million input tokens whether a prompt arrives in English or in Russian. What changes between the two languages is how many tokens the same paragraph becomes, and the bill is that count multiplied by the price.]]></description><content:encoded><![CDATA[
  <figure id="BWlF" class="m_original">
    <img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/ftmahlzyyccpqoqglblg.png" width="1600" />
  </figure>
  <p id="utBi">Claude Opus 5 costs $5 per million input tokens whether a prompt arrives in English or in Russian. What changes between the two languages is how many tokens the same paragraph becomes, and the bill is that count multiplied by the price.</p>
  <p id="VRTS">A recent test on <a href="https://habr.com/ru/articles/1089148/" target="_blank">Habr</a> measured the gap across 11 tokenizers. Its numbers line up with research going back to 2023: on most popular models, a language other than English costs more to send, fills the context window sooner and reaches rate limits with less text.</p>
  <h2 id="8uZR">How the test worked</h2>
  <p id="0FPw">A token is the chunk of text a model reads and bills by, usually a piece of a word. The tokenizer is the component that cuts text into tokens with a vocabulary fixed before the model is trained.</p>
  <p id="DErg">The author, Habr user donseo, wrote 4 parallel text pairs in Russian and English: a technical article, a support chat, a business notice and Python code with comments. The Russian side holds about 2,500 characters. Open tokenizers ran locally through <code>tiktoken</code> and Hugging Face. For closed models, the script sent each text with a short fixed prompt and the prompt alone, then subtracted the two token counts the API reported. The code and the corpus are public on <a href="https://github.com/donseo-info/ru-token-tax" target="_blank">GitHub</a> under MIT.</p>
  <h2 id="5O1I">Russian against English, same text</h2>
  <pre id="QgLZ">Model                      RU/EN tokens   Russian vs English
YandexGPT 5 Lite           0.91x          -9%
GigaChat 3                 0.96x          -4%
Grok 4.7                   1.18x          +18%
GPT-5.x / GPT-6 (o200k)    1.19x          +19%
Gemini 3.7 Flash           1.20x          +20%
GLM-5.3                    1.23x          +23%
DeepSeek V3/V4             1.38x          +38%
Qwen 3                     1.52x          +52%
Kimi K3                    1.86x          +86%
GPT-4 (cl100k)             2.09x          +109%
Claude Opus 5              2.96x          +196%</pre>
  <p id="oaZE">Claude Opus 5 sits far from the rest at <strong>2.96x</strong>. Its technical article took 327 tokens in English and 872 in Russian. Even its English is expensive: 3.25 characters per token, against 5.07 for the OpenAI o200k tokenizer behind GPT-5.x. In Russian it falls to 1.03 characters per token, close to one token per letter.</p>
  <p id="aOMN">The Python sample showed the smallest gap on Claude Opus 5, 424 tokens in Russian against 231 in English. Prose is where the premium bites hardest.</p>
  <p id="XiuJ">At the same per-token price, Russian text on Claude Opus 5 costs about 3.9x what it costs on GPT-5.5, from the tokenizer alone. That 3.9x compares two models on the same Russian text. Reposts of the table often present it as Russian against English, which overstates the language gap and hides the model gap.</p>
  <h2 id="EPks">Why the tokenizer favors English</h2>
  <p id="SLGg">A tokenizer builds its vocabulary from a training corpus, and frequent sequences in that corpus get their own tokens. Petrov et al. found that &quot;tokenizers are heavily influenced by the biases of the corpus source&quot;. Common English words become one token each. A Cyrillic word breaks into fragments, and in GPT-4&#x27;s cl100k_base tokenizer some Cyrillic letters take two tokens on their own.</p>
  <p id="7b5O">None of this depends on how well the model understands the language. The same paper puts it plainly: &quot;the unequal treatment of languages arises at the tokenization stage, well before the language model sees any data at all.&quot;</p>
  <p id="deps">Anthropic&#x27;s pricing page says the same thing in vendor terms. One token is &quot;approximately 4 characters or 0.75 words in English&quot;, and &quot;the exact count varies by language and content type&quot;. It also notes that Claude 4.7 and later models use a newer tokenizer that &quot;produces approximately 30% more tokens for the same text&quot;, a change that touches every language.</p>
  <h2 id="Hnp4">What extra tokens cost</h2>
  <p id="UNWz">Price, context and rate limits in an LLM API are all counted in tokens.</p>
  <ul id="jK5v">
    <li id="qa7B">Price. Anthropic bills per million tokens, $5 input and $25 output on Claude Opus 5. A prompt that becomes 2.96x the tokens costs 2.96x as much to send.</li>
    <li id="2HJm">Answers. Output is tokenized the same way, and on Claude Opus 5 an output token costs 5x an input token. A long answer in Russian carries the premium at the higher rate.</li>
    <li id="MWzQ">Context window. The window is a token budget. At 2.96x, it holds about a third as much Russian text as English, so long documents and chat histories run out of room sooner.</li>
    <li id="Xtqg">Rate limits. Anthropic measures them in input and output tokens per minute. A team working in Russian reaches the cap with less work done. 💸</li>
  </ul>
  <h2 id="FGft">Not a Russian problem</h2>
  <p id="ifNi">Petrov et al. measured parallel text in many languages for NeurIPS 2023. The tokenizer behind ChatGPT and GPT-4 used about 1.6x the tokens of English for Italian, 2.6x for Bulgarian and 3x for Arabic. For some languages the gap reached 15x. Their Russian figure on cl100k_base was 2.49x, on a different corpus from the Habr test. The authors&#x27; conclusion about money is direct: &quot;the tokenization premiums discussed in Section 4 directly map to cost premiums.&quot;</p>
  <p id="0ARx">Newer tokenizers have narrowed the gap for some vendors. OpenAI&#x27;s o200k came in at 1.19x for Russian in the Habr test. The two studies used different texts, so this shows a direction rather than a precise trend.</p>
  <p id="9VH5">Price is not the only cost. Ahia et al. at EMNLP 2023 found that speakers of many supported languages &quot;are overcharged while obtaining poorer results&quot;. Anthropic publishes scores relative to English for Claude Sonnet 4.5: Spanish at 98.2%, Chinese at 96.9%, Hindi at 96.7%, Swahili at 91.1% and Yoruba at 79.7%. Major languages stay close to English, smaller ones fall behind. Russian is not in that table. Wendler et al. add a hint about why English stays the default: inside the models they studied, &quot;the abstract &#x27;concept space&#x27; lies closer to English than to other languages&quot;.</p>
  <h2 id="kOHX">Where the rule breaks</h2>
  <p id="7Lmj">YandexGPT 5 Lite and GigaChat 3 use fewer tokens for Russian than for English. They are the two models in the test built for Russian, and their tokenizers encode Russian at least as compactly as English. That is the same mechanism from the other side: the language a tokenizer saw most is the language it encodes cheaply.</p>
  <p id="KsDj">The test has known limits. Its author lists them: the corpus is small, and the coefficients may move by about 5% on other texts. Closed-model counts came through third-party API gateways. The GigaChat figure uses the open GigaChat 3 tokenizer, which may differ from the production model. Claude&#x27;s tokenizer is not public, so its number rests on API usage counts.</p>
  <h2 id="JACl">What to do with it</h2>
  <ul id="Bnsv">
    <li id="5Lw3">Write system prompts, instructions and tool schemas in English when the model tokenizes your language poorly. The Habr author recommends exactly this.</li>
    <li id="Fg81">Compare models by the cost of a typical request on your own text, not by the price per million tokens.</li>
    <li id="4Sl2">Count before choosing. Anthropic&#x27;s token counting endpoint is free, and <code>tiktoken</code> covers OpenAI&#x27;s tokenizers.</li>
    <li id="ZUkV">For workloads that must stay in Russian, the Russian-built models start with a tokenizer advantage. 🧭</li>
  </ul>
  <p id="j4Ry"><strong>A price per token is only comparable between languages when the token counts are.</strong></p>
  <p id="DKcZ">For anyone paying per token, the language of the prompt is a pricing decision, and on Claude Opus 5 a Russian prompt buys about a third of the text the same money buys in English.</p>
  <h2 id="DCaN">References</h2>
  <ul id="CuF5">
    <li id="fG02"><a href="https://habr.com/ru/articles/1089148/" target="_blank">Habr: Russian text in Claude Opus 5 costs 3.9x more than in GPT-5.5 at the same per-token price (in Russian)</a></li>
    <li id="0uwm"><a href="https://github.com/donseo-info/ru-token-tax" target="_blank">ru-token-tax: code and corpus of the Habr test</a></li>
    <li id="fVAO"><a href="https://arxiv.org/abs/2305.15425" target="_blank">Petrov et al., Language Model Tokenizers Introduce Unfairness Between Languages</a></li>
    <li id="hGdQ"><a href="https://aclanthology.org/2023.emnlp-main.614/" target="_blank">Ahia et al., Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Models</a></li>
    <li id="ECWN"><a href="https://arxiv.org/abs/2402.10588" target="_blank">Wendler et al., Do Llamas Work in English? On the Latent Language of Multilingual Transformers</a></li>
    <li id="i0Ed"><a href="https://platform.claude.com/docs/en/about-claude/pricing" target="_blank">Anthropic: Pricing</a></li>
    <li id="1pDe"><a href="https://platform.claude.com/docs/en/build-with-claude/multilingual-support" target="_blank">Anthropic: Multilingual support</a></li>
    <li id="yPw7"><a href="https://platform.claude.com/docs/en/api/rate-limits" target="_blank">Anthropic: Rate limits</a></li>
    <li id="Luxv"><a href="https://platform.claude.com/docs/en/build-with-claude/token-counting" target="_blank">Anthropic: Token counting</a></li>
  </ul>
  <hr />
  <p id="ksGC">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/fable-advisor-opus-5-5-claude-code</guid><link>https://teletype.in/@khaskydev/fable-advisor-opus-5-5-claude-code?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/fable-advisor-opus-5-5-claude-code?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Fable becomes the advisor to Opus 5.5 in Claude Code</title><pubDate>Sat, 03 Oct 2026 02:23:23 GMT</pubDate><media:content medium="image" url="https://img2.teletype.in/files/1f/23/1f2398ad-1be1-4a5a-917f-e63c0d511044.png"></media:content><description><![CDATA[<img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/aa8ajty652g425kq8uaa.png"></img>With Opus 5.5 as the main model in a Claude Code session, a Fable quota can still go to work without Fable taking over the session. The /advisor fable command connects Fable as an advisor that Opus consults on its own. Opus keeps writing the code.]]></description><content:encoded><![CDATA[
  <figure id="2KOT" class="m_original">
    <img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/aa8ajty652g425kq8uaa.png" width="1600" />
  </figure>
  <p id="WS1c">With Opus 5.5 as the main model in a Claude Code session, a Fable quota can still go to work without Fable taking over the session. The <code>/advisor fable</code> command connects Fable as an advisor that Opus consults on its own. Opus keeps writing the code.</p>
  <h2 id="wHIU">What the advisor does</h2>
  <p id="C7pe">The advisor is a second model attached to the session, and Opus decides when to call it. It tends to ask before committing to an approach, when the same error keeps coming back, and before declaring a task done. That timing is model-driven rather than rule-based.</p>
  <p id="yRDE">Each call hands Fable the full conversation, with every tool call and its result. Fable returns guidance, and Opus applies it before it continues. With Opus 5 or Opus 5.5 as the main model, Claude Code accepts Fable or Opus 5 and later in the advisor role.</p>
  <h2 id="hHc5">Turning it on</h2>
  <p id="QUf6">On some plans Fable bills to usage credits, and Claude Code will not apply Fable as the advisor until a one-time consent is given. The consent comes from <code>/model fable</code>, and only after it does <code>/advisor fable</code> take effect:</p>
  <pre id="sUoy">/model fable
/advisor fable</pre>
  <p id="mWyI">Accepting the consent also makes Fable the selected main model. Switch back to Opus 5.5 with <code>/model</code> so the pairing stays Opus at the keyboard and Fable on call.</p>
  <h2 id="VqcT">Why it costs less</h2>
  <p id="GZdT">Fable is called at decision points rather than on every turn. A session paired this way typically costs less than one that runs the stronger model throughout. Each call still bills Fable&#x27;s tokens at Fable&#x27;s rates, on top of what Opus uses. 🧾</p>
  <p id="Mxgc">Turning the advisor on or off mid-session does not invalidate the main model&#x27;s prompt cache, unlike a model switch. Fable&#x27;s own read is not cached, though. Every call processes the full transcript anew.</p>
  <h2 id="rWAQ">Where it runs</h2>
  <p id="BJFu">The advisor is a server-side tool, so it needs the Anthropic API. It is not available on Amazon Bedrock, Claude Platform on AWS, Google Cloud&#x27;s Agent Platform or Microsoft Foundry.</p>
  <h2 id="cfQr">When the advisor stays silent</h2>
  <p id="8sdx">Two environment variables keep it off. <code>CLAUDE_CODE_DISABLE_ADVISOR_TOOL</code> disables the tool entirely, and the <code>/advisor</code> command disappears with it. <code>DISABLE_TELEMETRY</code> stops Claude Code from fetching the feature flag that turns the advisor on.</p>
  <p id="c9ks"><strong>When no advice ever arrives, check those two variables first.</strong></p>
  <h2 id="Hvxr">A setup that uses it</h2>
  <p id="DRUG">Opus 5.5 on high effort runs the session. Three subagents on medium effort take the routine work: one reads the code, one makes edits with tests, one looks up documentation. Fable waits on the side and comes in when the work needs a second look. 👀</p>
  <p id="pPTq">Subagents inherit the configured advisor and run the same pairing check against their own model. Each one can set its own <code>effort</code> in its frontmatter.</p>
  <p id="SV4l">For an Opus 5.5 session on the Anthropic API, the advisor is the cheaper way to put Fable&#x27;s judgment into the work, and keeping Fable as the main model now needs a reason beyond the occasional hard call.</p>
  <h2 id="P7jW">References</h2>
  <ul id="VXk1">
    <li id="7TPD">Advisor tool in Claude Code: <a href="https://code.claude.com/docs/en/advisor" target="_blank">https://code.claude.com/docs/en/advisor</a></li>
    <li id="Nnl3">Subagents and their frontmatter: <a href="https://code.claude.com/docs/en/sub-agents" target="_blank">https://code.claude.com/docs/en/sub-agents</a></li>
  </ul>
  <hr />
  <p id="cHQ7">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/rejudge-replaces-self-review-with-3-models</guid><link>https://teletype.in/@khaskydev/rejudge-replaces-self-review-with-3-models?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/rejudge-replaces-self-review-with-3-models?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Rejudge replaces self-review with 3 independent models and a judge</title><pubDate>Thu, 01 Oct 2026 22:20:11 GMT</pubDate><media:content medium="image" url="https://img3.teletype.in/files/ae/87/ae876bd3-a783-4f86-914d-0a192968f4d0.png"></media:content><category>AI</category><description><![CDATA[<img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/hjids1gq694pcgl4d56k.png"></img>I have a hard time trusting an AI coding agent to review code it just wrote.]]></description><content:encoded><![CDATA[
  <figure id="oMaJ" class="m_original">
    <img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/hjids1gq694pcgl4d56k.png" width="1600" />
  </figure>
  <p id="wASO">I have a hard time trusting an AI coding agent to review code it just wrote.</p>
  <p id="KoLi">The model already made the decision that the implementation was good enough to produce.</p>
  <p id="1M6C">Then we ask it to inspect that same implementation using many of the same learned habits and assumptions.</p>
  <p id="GffT">Sometimes it catches mistakes. Sometimes it revalidates them. A fresh session helps. A different model helps more.</p>
  <p id="5eNU">But once different models disagree, somebody still has to decide which review to trust.</p>
  <p id="Jg3H"><a href="https://github.com/syabro/rejudge" target="_blank">Rejudge</a> makes that disagreement-resolution step explicit.</p>
  <h2 id="BLyO">The architecture</h2>
  <p id="jR26">The same review request goes to three reviewers:</p>
  <pre id="hySj">reviewer A
reviewer B
reviewer C</pre>
  <p id="lJZ0">Each works in an isolated context. They do not see each other&#x27;s reasoning, tools, or conclusions. After all three finish, a separate judge receives their reports. The judge can ask follow-up questions when the panel disagrees. Then it writes one final answer.</p>
  <pre id="lYan">same question
    |
    +--&gt; reviewer A
    +--&gt; reviewer B
    +--&gt; reviewer C
             |
           judge
             |
        final answer</pre>
  <p id="0In6">Independence happens before collaboration.</p>
  <h2 id="8ZNJ">How reviewers inspect the code</h2>
  <p id="J7Pp">By default, reviewer tools include:</p>
  <pre id="JYLG">read
grep
find
ls
git_diff
web_search (if the host provides one)</pre>
  <p id="wQcj">The normal reviewers do not get:</p>
  <pre id="PeFL">edit
write
bash</pre>
  <p id="jnfL">That is a sensible default for code review. 🔒</p>
  <h2 id="0Obg">How judge makes a decision</h2>
  <p id="r1Yi">The judge gets no workspace access. It sees the three reports.</p>
  <p id="SxhJ">If those reports conflict, it can call <code>ask_panel</code> and request clarification.</p>
  <p id="iCoG">That means the judge is not a hidden fourth reviewer. Its role is:</p>
  <pre id="bNjA">compare
challenge
adjudicate
synthesize</pre>
  <h2 id="aDYV">Practical usage</h2>
  <p id="q9cV">Install (it needs Node.js 22.19.0 or newer):</p>
  <pre id="y4eS">npm install -g rejudge</pre>
  <p id="1AWX">Review a diff:</p>
  <pre id="en73">git diff | rejudge &quot;review this change&quot;</pre>
  <p id="GpRX">Ask a targeted question:</p>
  <pre id="Ax8u">rejudge &quot;does this migration need a lock?&quot;</pre>
  <p id="gsfC">Resume later:</p>
  <pre id="Of0N">rejudge --resume &lt;run-id&gt; &quot;what about the rollback path?&quot;</pre>
  <p id="nxoG">The answer goes to stdout, while progress, the config in use and the run ID go to stderr, so redirecting stdout to a file leaves only the answer. A resumed run reopens the same sessions, and the new question goes to the judge first. The reviewers hear it only if the judge calls <code>ask_panel</code>.</p>
  <h2 id="7tom">Coding-agent integration</h2>
  <p id="3GGO">Rejudge supports:</p>
  <ul id="Rgej">
    <li id="6SWO">CLI</li>
    <li id="SUTr">native Pi tool</li>
    <li id="3NGQ">Agent Skill for coding agents outside Pi</li>
  </ul>
  <p id="VzT8">Agent Skill install:</p>
  <pre id="f1r2">npx skills add syabro/rejudge -g -y</pre>
  <p id="EAnS">The skills are a separate copy, so the README says to refresh them after each Rejudge release with <code>npx skills update -g -y</code>. Inside Pi, the extension is one more line:</p>
  <pre id="4nEd">pi install &quot;$(npm root -g)/rejudge&quot;</pre>
  <p id="FSyP">Rejudge runs on Pi and reads its provider settings, so a key Pi already accepts works here too.</p>
  <h2 id="iBdX">The models are configurable</h2>
  <p id="J3AZ">Rejudge is not tied to one fixed provider combination. The config has a reviewer list and a separate judge model, with two reviewers as the minimum. Each model also gets a reasoning level from:</p>
  <pre id="NZXF">minimal
low
medium
high
xhigh</pre>
  <p id="T6zI">The global file lives at <code>~/.config/rejudge/config.json</code>, and a <code>.rejudge/config.json</code> in the project wins over it. That makes it possible to build a genuinely mixed panel.</p>
  <h2 id="0lJa">There is an unsafe mode</h2>
  <p id="b4CI"><code>--unsafe</code> / <code>--full</code> gives reviewers:</p>
  <pre id="SNbQ">edit
write
bash</pre>
  <p id="XvxV">The docs explicitly say this is <strong>not a sandbox</strong>. The judge still gets only <code>ask_panel</code>. For review-only work, I would stay read-only.</p>
  <h2 id="XIdp">Multiple providers mean multiple privacy boundaries</h2>
  <p id="efNY">Every reviewer model sees the request. Anything a reviewer reads becomes part of that provider session.</p>
  <p id="E9fk">The judge does not inspect the workspace directly, but reviewer reports may quote code.</p>
  <p id="V3Sy">Read-only tools stop local changes. They do not keep file contents private, and the README warns that instructions hidden in a request or in a file a reviewer opens can steer what it reads and reports.</p>
  <p id="7XAA">Runs also leave records. Sessions are written to <code>${TMPDIR}/rejudge/runs/&lt;run-id&gt;/</code> while a run executes, and cleanup after roughly 24 hours is best-effort. The <code>debugLog</code> option is off by default, and turning it on writes full model thinking to <code>.rejudge/logs/</code>.</p>
  <p id="jBBg">For proprietary repositories, this needs to be acceptable before running the panel.</p>
  <h2 id="VoiO">The cost</h2>
  <p id="qS5M">A fresh review starts with:</p>
  <pre id="hK38">3 reviewer calls
+ 1 judge call</pre>
  <p id="gd2r">Then add tool loops, retries, recovery, and judge follow-ups. Rejudge has no spending cap of its own. This is not a free accuracy multiplier. It is a compute tradeoff. 💸</p>
  <h2 id="Eakc">Where I would use it</h2>
  <p id="Zx7M">I would reserve it for changes where mistakes are expensive:</p>
  <pre id="aPTd">database migrations
locking/concurrency
authentication
authorization
permissions
security-sensitive code
data deletion
rollback logic
complex refactors</pre>
  <p id="5UfZ">Rejudge is a structured independent second opinion. For consequential code, that can be worth the extra cost.</p>
  <h2 id="McBB">References</h2>
  <ul id="TWKE">
    <li id="nHxZ"><a href="https://github.com/syabro/rejudge" target="_blank">Rejudge on GitHub</a>: README, install, config and the privacy and cost notes</li>
    <li id="2CA8"><a href="https://rejudge.syabro.com" target="_blank">Rejudge site</a>: overview and demo</li>
  </ul>
  <hr />
  <p id="xhWL">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/gemini-4-argon-beats-gpt-6-astra-and-claude</guid><link>https://teletype.in/@khaskydev/gemini-4-argon-beats-gpt-6-astra-and-claude?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/gemini-4-argon-beats-gpt-6-astra-and-claude?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Gemini 4 Argon beats GPT-6 Astra and Claude on most benchmarks</title><pubDate>Thu, 01 Oct 2026 06:15:00 GMT</pubDate><media:content medium="image" url="https://img3.teletype.in/files/e5/e1/e5e15fce-2f22-4d06-9b9e-684b3fd7dbaf.png"></media:content><description><![CDATA[<img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/102mok8kdl4g93wyv8gp.png"></img>Google DeepMind has announced Gemini 4 Argon, a frontier model aimed at long software engineering jobs, legal and finance work, and cyber defense. In Google's own evaluation it beats GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on most of the benchmarks it reports. Almost nobody can use it yet.]]></description><content:encoded><![CDATA[
  <figure id="9C83" class="m_original">
    <img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/102mok8kdl4g93wyv8gp.png" width="1600" />
  </figure>
  <p id="zNgy">Google DeepMind has announced Gemini 4 Argon, a frontier model aimed at long software engineering jobs, legal and finance work, and cyber defense. In Google&#x27;s own evaluation it beats GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on most of the benchmarks it reports. Almost nobody can use it yet.</p>
  <h2 id="GctD">Where Argon comes first</h2>
  <p id="qF4w">The comparison table covers 19 rows across knowledge work, agentic coding, ML engineering, science and math, long context, computer use, multimodal understanding and cybersecurity. Argon finishes first or tied for first on 14 of them.</p>
  <p id="KMDB">The number Google leads with is DeepSWE v1.1, a set of real-world software engineering tasks. Argon scores 77.9% there, against 74.2% for Opus 5.5, 74.1% for Astra and 67.4% for Fable 5.1. The widest gap sits on Harvey&#x27;s Legal Agent Benchmark, where Argon reaches 19.6% and none of the other three gets past 6.7%. 🎯</p>
  <pre id="QeEh">Benchmark                          Argon   Astra   Fable 5.1  Opus 5.5
Vals Index                         68.9    63.1    65.8       67.0
AutomationBench                    51.3    41.4    31.4       42.5
Vals Finance Agent v2              65.4    53.5    58.9       58.6
Harvey&#x27;s Legal Agent Benchmark     19.6     5.4     6.7        3.8
DeepSWE v1.1                       77.9    74.1    67.4       74.2
FrontierSWE v2                     55.0    65.5    56.3       62.3
Vibe Code Bench                    91.9    89.6    90.3       90.3
Terminal-bench 4.0                 57.4    58.2    57.9       66.4
PostTrainBench                     45.3    44.3    40.2       49.3
Terminal-Bench Science 0.1         57.6    68.1    52.6       63.3
LABBench 2                         88.8    85.4    68.6       73.1
RiemannBench                       76.0    72.0    65.6       69.6
GraphWalks BFS, up to 128k         99.7    98.7    91.4       90.6
GraphWalks BFS, 256k to 1M         84.2    71.8    65.0       66.8
Agent&#x27;s Last Exam (pass rate)      39.5    34.2    -          38.2
OSWorld-2.0 (offline, partial)     69.2    72.6    -          -
Chartography                       71.6    71.0    46.2       66.3
LVBench                            91.7    87.5    79.7       83.7
CWE-bench v1                       68.0    68.0    58.0       67.0</pre>
  <p id="g0tJ">All values are percentages from Google&#x27;s table. GraphWalks reports F1, and OSWorld-2.0 uses an offline subset with partial scores.</p>
  <h2 id="zsqD">Where it does not</h2>
  <p id="bPVH">Five rows go to someone else, and one is a tie. Astra keeps FrontierSWE v2 with 65.5% against Argon&#x27;s 55.0%, Terminal-Bench Science 0.1 with 68.1% against 57.6%, and the OSWorld-2.0 offline subset with 72.6% against 69.2%. Opus 5.5 keeps Terminal-bench 4.0 with 66.4% against 57.4%, and PostTrainBench with 49.3% against 45.3%. On CWE-bench v1 Argon and Astra share first place at 68.0%.</p>
  <p id="anbQ">Every figure here comes from Google&#x27;s own runs, and the methodology page is the place to check how the rival models were set up.</p>
  <h2 id="yAt3">One million tokens in one answer</h2>
  <p id="g7lb">Argon can write up to 1M output tokens in a single response. The previous limit was 64K, so one answer can now run about 15 times longer. Google ties this to the work the model targets: long, multi-step jobs where the reasoning has to hold across the whole run.</p>
  <h2 id="xajL">The price now and later</h2>
  <p id="KIpt">The introductory API price is $2 per 1M input tokens and $10 per 1M output tokens, and cached input costs 95% less than regular input. When the introductory period ends, the rate becomes $4 and $20.</p>
  <p id="mbLH">GPT-6 Astra lists $10 per 1M input tokens and $50 per 1M output tokens at its standard short-context rate. That puts Argon at 5x cheaper than Astra on both sides today and 2.5x cheaper after the increase. The comparison is token price against token price, and a model that writes longer answers can still cost more per finished task. 💸</p>
  <h2 id="OKnW">What Google already used it for</h2>
  <p id="KFWs">Inside Google, Argon has been working on real projects before release. Quantum computing researchers used it to optimize spacetime resources, counted as qubits times gates, and in one case it beat the published baseline by 40% within minutes.</p>
  <p id="L1EI">A team of Argon agents read fleet-wide profiling telemetry across Google&#x27;s data centers and applied memory optimizations on its own, freeing more than 300 TiB of memory.</p>
  <p id="waLj">Argon agents are also moving C and C++ codebases to Rust. Early reports put that at tens of thousands of lines, which holds for core libraries such as re2 and libgav1, while the largest job, the Fuchsia Zircon kernel, runs past 800K lines. The libgav1 result is a memory-safe video decoder with identical output that runs 2.7x faster than the earlier Rust port.</p>
  <h2 id="aZud">Who can use it</h2>
  <p id="sPou">Right now Argon is going to a set of trusted cyber defenders through Google&#x27;s Fairwind Program. Paid API customers and Google AI Ultra subscribers come next, followed by developers, enterprises and consumers more broadly. Google has given no date.</p>
  <p id="b0kl">Until Argon reaches the paid API, the table and the $2 price are something to plan a test around, and nobody outside Fairwind can yet check whether those 14 first places hold up.</p>
  <h2 id="UHRO">References</h2>
  <ul id="tc5e">
    <li id="KL6l"><a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/" target="_blank">Google: Gemini 4 Argon, our next era of frontier intelligence</a></li>
    <li id="niuy"><a href="https://deepmind.google/models/evals-methodology/gemini-4-argon" target="_blank">Google DeepMind: Gemini 4 Argon evaluation methodology</a></li>
    <li id="Vpbx"><a href="https://developers.openai.com/api/docs/pricing" target="_blank">OpenAI API pricing</a></li>
  </ul>
  <hr />
  <p id="nMJn">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/free-email-forwarding-cloudflare-email-routing</guid><link>https://teletype.in/@khaskydev/free-email-forwarding-cloudflare-email-routing?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/free-email-forwarding-cloudflare-email-routing?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Free email forwarding for your personal domain with Cloudflare Email Routing</title><pubDate>Tue, 29 Sep 2026 07:44:51 GMT</pubDate><media:content medium="image" url="https://img1.teletype.in/files/0c/24/0c248e58-7386-4196-aec1-2652aeba2e28.png"></media:content><category>Development</category><description><![CDATA[<img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/64mh1ohn4kmvsim7bgde.png"></img>Forward any address at your domain, such as hello@example.com, to an external inbox. For a public address that only has to receive mail, I would put it on Cloudflare Email Routing: the messages land in the inbox I already read, and forwarding to a verified destination costs nothing on any plan. 💸]]></description><content:encoded><![CDATA[
  <figure id="0vx5" class="m_original">
    <img src="https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/64mh1ohn4kmvsim7bgde.png" width="1600" />
  </figure>
  <p id="kXpp">Forward any address at your domain, such as hello@example.com, to an external inbox. For a public address that only has to receive mail, I would put it on Cloudflare Email Routing: the messages land in the inbox I already read, and forwarding to a verified destination costs nothing on any plan. 💸</p>
  <p id="Ir9h">Prerequisites: the domain is on Cloudflare DNS (nameservers delegated, DNS Setup = Full), and it has no MX records from another mail host. Email Routing owns the MX set.</p>
  <h2 id="Rf5k">1. Verify the destination address</h2>
  <p id="gKLr">Compute &gt; Email Service &gt; Email Routing &gt; Destination Addresses &gt; Add address. Enter the external inbox (for example, <code>you@gmail.com</code>). Cloudflare emails a confirmation link, and you click it.</p>
  <p id="SrXw">The row must read Verified. Until it does, any rule that points to the address stays disabled.</p>
  <p id="PocW">Destination addresses are account-wide: once verified, any zone in the account can target it.</p>
  <h2 id="0pPj">2. Enable routing and add the DNS records</h2>
  <p id="afeo">Open the zone&#x27;s Email Routing &gt; Settings &gt; DNS records &gt; Add missing records. Cloudflare writes these records and locks them:</p>
  <pre id="IZkY">Type  Name                             Value
MX    example.com                      route1.mx.cloudflare.net (+ route2, route3)
TXT   example.com                      v=spf1 include:_spf.mx.cloudflare.net ~all
TXT   cf2024-1._domainkey.example.com  v=DKIM1; h=sha256; k=rsa; p=...</pre>
  <p id="g1lk">A locked record cannot be edited or deleted from the DNS page until you unlock it in the Email Routing settings.</p>
  <p id="WQvj">Keep one SPF TXT record per name. If the domain already has one, the merge is yours to do: Cloudflare&#x27;s example combines both includes in a single record, <code>v=spf1 include:_spf.mx.cloudflare.net include:_spf.google.com ~all</code>.</p>
  <h2 id="RfEb">3. Create the routing rule</h2>
  <p id="2n1u">Email Routing &gt; Routing rules &gt; Create routing rule:</p>
  <ul id="gV6D">
    <li id="pPCo">Email pattern: <code>hello</code> @ <code>example.com</code></li>
    <li id="eqdT">Action: Send to an email</li>
    <li id="fS4j">Destination: the verified address</li>
  </ul>
  <p id="TUTi">Save. The rule must show Active.</p>
  <p id="sFS5">The pattern takes any local part, not only <code>hello</code>: <code>sales</code>, <code>support</code> or your own name work the same way, one rule per address.</p>
  <h2 id="kU7y">4. Decide the catch-all</h2>
  <p id="VfEc">The catch-all in Routing rules defines what happens to every other address at the domain. Enabled with Send to an email, it forwards all of them, typos such as <code>ifno@example.com</code> included. Enabled with Drop, it discards them silently.</p>
  <p id="uONC">I would start with it off and turn on forwarding only if people keep misspelling the real address.</p>
  <h2 id="BSf8">5. Verify</h2>
  <pre id="fvEO"># MX must answer from the Cloudflare route hosts
curl -sH &#x27;accept: application/dns-json&#x27; \
  &#x27;https://cloudflare-dns.com/dns-query?name=example.com&amp;type=MX&#x27;

# SPF
curl -sH &#x27;accept: application/dns-json&#x27; \
  &#x27;https://cloudflare-dns.com/dns-query?name=example.com&amp;type=TXT&#x27;</pre>
  <p id="3kyL">Then send a real message to hello@example.com and confirm it lands in the destination inbox. 📬</p>
  <p id="Z0qq">Email Routing &gt; Activity Log shows the decision for each message: Forwarded, Dropped, Rejected or Delivery failed.</p>
  <h2 id="O7sH">Notes</h2>
  <ul id="2Fna">
    <li id="fUUj">Forwarding only. A routing rule receives mail and does not send it, so a reply from the destination inbox goes out through that provider, from its address. Sending from the domain is a separate product, Cloudflare Email Sending, which is still in beta.</li>
    <li id="GT6g">Negative DNS answers stay cached for the zone&#x27;s SOA minimum TTL, 1800 seconds (30 minutes) by default on Cloudflare. If you queried MX before the records existed, your resolver may keep returning empty. Query a different resolver or go over DoH to confirm.</li>
  </ul>
  <p id="Opr7">Five steps give a domain working addresses with no mailbox bill, and I would pay for a mailbox only on the day replies have to come from the domain itself.</p>
  <h2 id="12Eg">References</h2>
  <ul id="Tv5S">
    <li id="vmON"><a href="https://developers.cloudflare.com/email-routing/" target="_blank">Email Routing overview and pricing</a></li>
    <li id="oRTN"><a href="https://developers.cloudflare.com/email-routing/get-started/enable-email-routing/" target="_blank">Enable Email Routing</a></li>
    <li id="YaU4"><a href="https://developers.cloudflare.com/email-routing/setup/email-routing-dns-records/" target="_blank">Email Routing DNS records</a></li>
    <li id="YOLc"><a href="https://developers.cloudflare.com/email-routing/setup/email-routing-addresses/" target="_blank">Destination addresses and routing rules</a></li>
    <li id="iX71"><a href="https://developers.cloudflare.com/1.1.1.1/encryption/dns-over-https/make-api-requests/dns-json/" target="_blank">DNS over HTTPS JSON queries</a></li>
  </ul>
  <hr />
  <p id="jLJO">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/80-open-source-pull-request-templates</guid><link>https://teletype.in/@khaskydev/80-open-source-pull-request-templates?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/80-open-source-pull-request-templates?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Inside 80 Pull Request Templates From Popular Open Source Projects</title><pubDate>Wed, 23 Sep 2026 21:07:50 GMT</pubDate><media:content medium="image" url="https://img4.teletype.in/files/30/9f/309fd116-e2c5-4be5-8809-043a1ec92036.png"></media:content><category>Development</category><description><![CDATA[<img src="https://img3.teletype.in/files/a0/29/a0294853-8809-407d-9e4d-53adc3c450eb.png"></img>A pull request template can ask for very little, and most of 80 well-known open source repositories do exactly that.]]></description><content:encoded><![CDATA[
  <figure id="D3XY" class="m_original">
    <img src="https://img3.teletype.in/files/a0/29/a0294853-8809-407d-9e4d-53adc3c450eb.png" width="1600" />
  </figure>
  <p id="LYgx">A pull request template can ask for very little, and most of <strong>80 well-known open source repositories</strong> do exactly that.</p>
  <p id="KsqO">I fetched every template straight from the GitHub, and the shape that turns up most often has exactly two prompts: an account of the change with its reason, and a description of how the author checked it.</p>
  <p id="hx5V">I read them the way a designer reads a form. Every field costs the person filling it in. Every field should also hand the reviewer something they could not get elsewhere, and judged that way, the boxes that restate CI fail first.</p>
  <h3 id="YF82">Templates that render as nothing</h3>
  <p id="0DMQ">CPython, curl, Deno, etcd, Laravel, Node, Rust, Spring Boot, Swift, Vite and VS Code all belong here.</p>
  <p id="eTt8">Their sizes vary a lot. etcd needs 7 lines and VS Code 8. Laravel uses 9, Vite 21 and Swift 22, while Spring Boot reaches 29 and Node 45. Length makes no difference to what the reviewer sees, which is nothing.</p>
  <p id="q4Qx">What they say while hidden is revealing. Rust’s comment explains its own <em>homu-ignore:start</em> and <em>homu-ignore:end</em> markers, which tell the merge bot to strip the text. CPython’s carries only title guidance, the <em>gh-NNNNNN:</em> prefix followed by a summary. Node carries the Developer’s Certificate of Origin 1.1 in full. Deno keeps a numbered list of rules. Swift describes the change it expects and how to trigger CI. curl carries nothing but an AI policy: “If you cannot understand or explain your work without using Artificial Intelligence (AI) then do not file here.”</p>
  <p id="wM2N">So these files are notices, closer to a sign on a door than to a form. They most often cover which branch to target, where security issues go and what not to send. A reviewer never sees them, and neither does anyone reading the history years later.</p>
  <p id="L29X">That design has one clear use. It reaches the author at the last moment before submission, when a wrong target branch is still cheap to fix.</p>
  <h3 id="9fJ8">Projects with no template</h3>
  <p id="Mby2">19 of the 80 repositories that answered ship no pull request template at all.</p>
  <blockquote id="0ojN">Vue core, webpack, Nuxt, React Router, Preact, Playwright,<br />Express, Fastify, TensorFlow, DuckDB, containerd, Mastodon,<br />OBS Studio, uBlock Origin, Kotlin, Ruby, PHP, LLVM, workerd</blockquote>
  <p id="VBzd">These are not the small ones, and they come from every domain the survey covered, from compilers and runtimes to databases, editors and web frameworks. Several maintain elaborate issue templates, so leaving the pull request box empty is a decision about where structure belongs.</p>
  <h3 id="Bwd1">The common template is two prompts long</h3>
  <p id="XnaL">The remaining templates, roughly fifty, put visible structure into a description. Their median is small, and the shape that turns up most is a pair of prompts: the change with its purpose, then how it was tested.</p>
  <blockquote id="8RNy">### What does this PR do?### How did you verify your code works?</blockquote>
  <p id="2tkW">That is Bun’s template, word for word, three lines in all. React, Solid, Tailwind, Neovim and Biome keep the same two ideas and add little. React’s version has four visible lines and ends on the only stated consequence anywhere in the sample: “If you leave this empty, your PR will very likely be closed.”</p>
  <h3 id="LuEf">Long checklists come with heavy traffic</h3>
  <p id="Xrnd">Above the two-prompt floor, projects reach for very different tools.</p>
  <p id="KRGU">Angular splits the description into current behavior and new behavior, adds a breaking-change question and asks the author to pick one of nine PR types. Kubernetes runs to 92 lines with seven headings, from special notes for the reviewer to an AI usage disclosure. Grafana asks three plain questions: what the feature is, why it is needed and who it is for.</p>
  <p id="JBcc">Some projects drop prose entirely. Babel and Symfony both use a question-and-answer table, and Symfony’s is read by a bot for labelling. Envoy uses colon-terminated fields, such as Commit Message, Risk Level and Release Notes. Go writes a plain-text instruction list with no headings at all.</p>
  <p id="flET">PyTorch goes another way and offers a directory of three choosable templates, for a docs typo, a fix for an issue and a preapproved change.</p>
  <h3 id="DF8N">Section frequency</h3>
  <blockquote id="FMIu">what changed and why — nearly every visible template<br />linked issue — about two thirds<br />testing or verification — about half<br />a checklist — about a third<br />AI authorship disclosure — about a third, the newest<br />release note slot — ten or so<br />breaking change callout — under ten<br />risk and rollback — three</blockquote>
  <p id="soQE">What changed and why is nearly universal. Small templates phrase it as one question.</p>
  <p id="82uI">The linked issue is the next most common ask. Most projects mention it in a comment. PyTorch, Keras, LangChain, Django, TypeScript and Axios treat it as mandatory, and the strict ones tend to attach a threat of closure, since a soft request gets ignored.</p>
  <h3 id="Ccls">Checklists</h3>
  <p id="qjmI">Checklists concentrate in high-volume consumer projects. The three heaviest checklist templates are Home Assistant at 119 lines, Transformers at 91 and Storybook at 86, and every one of them receives a stream of contributions from people who will submit once. The boxes screen that stream before a human looks. Home Assistant even has a box asserting the author reviewed two other open pull requests.</p>
  <p id="GSUF">Checklists travel well beyond consumer apps. Angular, Svelte, Django, Rails, Vitest and pandas all carry one. Among consumer apps, Signal-Android asks the author to list the devices and Android versions they tested on. Element asks for guidelines, tests, screenshots and the CLA, plus a pledge not to force push again. Immich and cal.com both pair their checklist with a testing section written in plain words.</p>
  <p id="ZiNb">Where contributors are regulars, the checklist disappears. Neovim, curl, Git and Bun have none.</p>
  <h3 id="D2A8">CI boxes</h3>
  <p id="pWzI">A box asking whether lint passed or tests pass is a claim nobody verifies. The pipeline already knows the answer, and it gates the merge regardless. The tick has no effect.</p>
  <p id="0lfu">What earns space is the thing a pipeline can’t see. Spark asks for copy-pasteable steps when a change was tested outside the regular unit tests. Storybook asks separately how the author tested and how a maintainer should reproduce it. Zed and Syncthing ask how a reviewer gets to the change at all.</p>
  <h3 id="A6aQ">Release notes</h3>
  <p id="oCtL">Moby, Terraform and Envoy carry a fenced release-note block that a changelog tool scrapes, as do Kubernetes, Prometheus and Zed. Grafana goes further and uses the pull request title itself to auto-generate its changelog. Without that tooling, a release-note section is dead weight.</p>
  <h3 id="P3hk">Risk and rollback</h3>
  <p id="eVb0">Only three templates ask about risk. Terraform’s asks which release a change targets, promises a revert within seven days if one is needed, and wants any change to security controls named. Envoy gives risk its own labelled line. The dotnet servicing template asks for customer impact, regression status and risk, because a merge there can go into a release customers already run. No library or frontend framework asks.</p>
  <h3 id="ePlE">Titles</h3>
  <p id="rxEs">Where a title becomes a durable record, the template polices it. CPython wants the issue number up front. Go wants a package prefix like <em>net/http:</em>, a short title and no Markdown. Deno asks for a conventional-commit prefix with worked examples. The title ends up as a commit subject, a backport tag or a changelog line.</p>
  <h3 id="dhMJ">AI disclosure</h3>
  <p id="k5W0">About a third of templates now ask about AI, and the wording differs every time. Django requires one of two exclusive boxes. pandas wants to know the exact assistant, its model version and its effort level. Caddy pre-fills “This PR is missing an assistance disclosure” under its heading, so silence ships as an admission. Transformers tells first-time contributors not to use code agents. Kubernetes writes a line addressed to the agent.</p>
  <h3 id="NB3y">Two missing prompts</h3>
  <p id="bgy2">Why the defect wasn’t caught earlier appears in no template. It’s the one thing a reviewer cannot reconstruct from the diff, because the diff shows the fix and never the path by which the bug slipped past every earlier review, test and release.</p>
  <p id="12UX">The second gap is smaller. What the change still doesn’t prove appears only obliquely, in dotnet servicing through customer impact and in Envoy through its risk field. Both come close to asking where the author’s confidence ends, and nothing else in the sample does even that much.</p>
  <h3 id="GwnW">What I’d put in a template</h3>
  <p id="F886">For a small project with regular contributors, I’d write two prompts and stop:</p>
  <p id="y3KL">- the change and its reason, opening on what changed rather than on what broke<br />- how it was verified, phrased so that restating the pipeline obviously doesn’t count</p>
  <p id="n6dE">A linked issue joins them only where work really lives in issues. Release notes wait for a generator. Risk questions come in the day a merge could hurt a production system or someone’s data, and not a day before 🎯</p>
  <h3 id="R5gQ">Limits of the evidence</h3>
  <p id="GjyS">Everything here comes from one snapshot of default branches. Templates change fast. The AI sections alone are young enough that a year ago most of them didn’t exist. I picked repositories by popularity, and popular projects run heavier templates than most.</p>
  <p id="TmBt">The survey records what templates ask and says nothing about whether contributors answer, which is the question that would settle which fields work.</p>
  <h3 id="S0XG">References</h3>
  <p id="lKbt">- GitHub Docs, creating a pull request template: <a href="https://docs.github.com/en/communities/using-templates-to-encourage-useful-issues-and-pull-requests/creating-a-pull-request-template-for-your-repository" target="_blank">https://docs.github.com/en/communities/using-templates-to-encourage-useful-issues-and-pull-requests/creating-a-pull-request-template-for-your-repository</a></p>
  <hr />
  <p id="czNp">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/vals-ai-august-2027-full-rsi</guid><link>https://teletype.in/@khaskydev/vals-ai-august-2027-full-rsi?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/vals-ai-august-2027-full-rsi?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Vals AI names August 2027 for full recursive self-improvement</title><pubDate>Tue, 22 Sep 2026 23:12:16 GMT</pubDate><description><![CDATA[Rayan Krishnan has put a month on the moment AI stops needing its researchers. The co-founder and CEO of the evaluation company Vals AI says August 2027.]]></description><content:encoded><![CDATA[
  <p id="hZhe">Rayan Krishnan has put a month on the moment AI stops needing its researchers. The co-founder and CEO of the evaluation company Vals AI says August 2027.</p>
  <h2 id="34j1">The forecast</h2>
  <p id="dJO0">He expects models to surpass the people who build them by then, and he defines the bar as full recursive self-improvement. Not a higher test score. A model that autonomously develops its own next version without a researcher taking part, the way GPT-6 would build GPT-7 or Claude its own successor. 🔄</p>
  <h2 id="ieNd">What the model would take over</h2>
  <pre id="e12S">today       AI takes part in building new models, people make the key decisions
full RSI    hypotheses -&gt; experiments -&gt; training the next model, all run by the model</pre>
  <p id="skNl">Today AI already takes part in building new models, and people still make the key decisions. Under full RSI the whole research cycle, from the hypotheses and the experiments to training the next model, moves to the model itself.</p>
  <p id="0t9A">Higher benchmark scores are not the claim. The claim is the whole research cycle running with no researcher in it.</p>
  <h2 id="9hNU">The date</h2>
  <p id="zyP9">August 2027 is Vals AI&#x27;s own forecast, not a confirmed date.</p>
  <h2 id="AnDJ">The measurement behind it</h2>
  <p id="wqJ6">Vals AI publishes an RSI Index that scores models on autonomous research tasks in language-model development. Its top entry, Claude Fable 5.1, scores 35.03% under fixed compute and time limits, and the company&#x27;s own note says those models run experiments well but still lack the judgment to identify higher-leverage directions. 🔎</p>
  <p id="4VOW">The forecast and that note come from the same company, which is the fact to hold onto.</p>
  <p id="jVf1"></p>
  <p id="HkbQ">The interview: <a href="https://www.youtube.com/watch?v=MpYAoufS588" target="_blank">https://www.youtube.com/watch?v=MpYAoufS588</a></p>
  <p id="tPHq">The RSI Index: <a href="https://www.vals.ai/benchmarks/rsi_index" target="_blank">https://www.vals.ai/benchmarks/rsi_index</a></p>
  <p id="xm1f">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/open-weight-gap-mozilla-measured</guid><link>https://teletype.in/@khaskydev/open-weight-gap-mozilla-measured?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/open-weight-gap-mozilla-measured?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>The Open-Weight Gap Mozilla Measured, and Why the Money Lags</title><pubDate>Tue, 22 Sep 2026 08:19:34 GMT</pubDate><description><![CDATA[The best open model is 5 points behind the closed frontier on Epoch's index. Mozilla's report is mostly about what that small number does and does not buy. 📄]]></description><content:encoded><![CDATA[
  <p id="EkrW">The best open model is 5 points behind the closed frontier on Epoch&#x27;s index. Mozilla&#x27;s report is mostly about what that small number does and does not buy. 📄</p>
  <h2 id="pJ5O">The gap is small and it resets</h2>
  <p id="WuRo">Kimi K3 scores 157 against 162 for Claude Fable 5 and Claude Opus 5, and Epoch converts its average gap of 8 points into about four months. Nothing accumulates. Every model on the chart shipped between June and August, so each release cycle redraws the margin from scratch rather than adding to a running total. In August 8 of the 10 most-used models on OpenRouter were open weights, 7 of them Chinese-built.</p>
  <h2 id="ebgt">The money is old and small</h2>
  <p id="BW6T">In the May to September 2025 window on OpenRouter, open models carried about 20% of model-layer usage and took about 4% of the revenue, with closed models charging about 6x more per call at roughly 90% capability parity.</p>
  <pre id="XdBr">OpenRouter, May-September 2025, model layer
usage      open ~20%    closed ~80%
revenue    open ~4%     closed ~96%</pre>
  <p id="iiEv">That split is the last one published, and the report says usage has moved since.</p>
  <h2 id="TXV8">Where deployments stop</h2>
  <p id="XZo5">79% of surveyed developers use open models, but 51% of open deployments reach production against 63% for closed, and the survey traces the gap to operational tooling and trust, with capability already clearing the bar. 🔧</p>
  <h2 id="6Spl">Gains from post-training</h2>
  <p id="TvsK">DeepSeek&#x27;s July 31 post-training pass added 10 points to V4-Flash on the Artificial Analysis index with the architecture unchanged, and the August 13 pass added 8 to V4-Pro. Neither involved a new pretraining run, which the report reads as a shorter shelf life for its own tables.</p>
  <h2 id="gu9e">Limits</h2>
  <p id="MvMu">OpenRouter measures routed traffic only, and the gap itself may be understated, since open models optimize harder for benchmarks and unpublished closed models are absent from the baseline.</p>
  <p id="PIz4"><a href="https://stateofopensource.ai" target="_blank">https://stateofopensource.ai</a></p>
  <p id="q2rR">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item><item><guid isPermaLink="true">https://teletype.in/@khaskydev/your-agent-does-not-see-the-asterisks</guid><link>https://teletype.in/@khaskydev/your-agent-does-not-see-the-asterisks?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev</link><comments>https://teletype.in/@khaskydev/your-agent-does-not-see-the-asterisks?utm_source=teletype&amp;utm_medium=feed_rss&amp;utm_campaign=khaskydev#comments</comments><dc:creator>khaskydev</dc:creator><title>Your Agent Does Not See the Asterisks in SKILL.md</title><pubDate>Mon, 21 Sep 2026 23:46:05 GMT</pubDate><media:content medium="image" url="https://img3.teletype.in/files/a3/4c/a34cac88-e215-460d-954f-32ecb18232b1.png"></media:content><category>AI</category><description><![CDATA[<img src="https://img1.teletype.in/files/cf/b1/cfb1a1ac-1489-446a-a642-4ed8eb783d24.png"></img>Bold is a device for a reader, and an instruction file has two readers with nothing in common: one has an eye that stops on a marked line, the other receives the same characters in the same order 🧠]]></description><content:encoded><![CDATA[
  <figure id="knf5" class="m_original">
    <img src="https://img1.teletype.in/files/cf/b1/cfb1a1ac-1489-446a-a642-4ed8eb783d24.png" width="1600" />
  </figure>
  <p id="HOIm">Bold is a device for a reader, and an instruction file has two readers with nothing in common: one has an eye that stops on a marked line, the other receives the same characters in the same order 🧠</p>
  <h2 id="rN3S">The mechanism, and what the research shows</h2>
  <p id="bKub">There is no bold channel. Nothing inside these models turns an emphasized span into extra weight, and no documentation for Claude Code, Codex or Gemini claims that anything does.</p>
  <p id="180E">Nothing does.</p>
  <p id="y2e5">To make a marked span behave the way a reader expects, the researchers had to leave the prompt entirely. PASTA reweights a small subset of attention heads at inference, changes no weights, and reports an average accuracy improvement of <strong>22% on LLAMA-7B</strong>.</p>
  <h2 id="Z3JK">What the vendors document, and what I counted</h2>
  <p id="7OhU">Anthropic has one sentence about emphasis and it carries both halves:</p>
  <blockquote id="Lbnl">If Claude keeps skipping one instruction, add emphasis such as &quot;IMPORTANT&quot; to that line alone. If you emphasize many lines, none of them stands out.</blockquote>
  <p id="aAtd">Then I counted the file I had been maintaining. 635 bold spans across one skill, 91 of them in a 264 line file, one every three lines, and stripping every asterisk saves 2,540 characters, which is 1% of the text, so the token argument was dead before I made it.</p>
  <h2 id="CXdx">The limits of all this</h2>
  <p id="kz6d">I should be straight about the other side of this. No one has run bold against no bold as an experiment, there is no measured penalty anywhere to point at, and bold is neither harmful nor inert.</p>
  <h2 id="K9Xv">The replacement</h2>
  <p id="2F94">Position first: if the agent needs something at step four, the sentence goes at step four. After that it is one or two hard words per file, NEVER or MUST, kept scarce enough to register when they appear, and then enforcement for anything that cannot tolerate a single exception, because prose asks and a hook decides.</p>
  <p id="3aTC">Only one of those three removes the question entirely 🔧</p>
  <ul id="CnbM">
    <li id="XPlj">Anthropic&#x27;s Claude Code best practices: <a href="https://code.claude.com/docs/en/best-practices" target="_blank">https://code.claude.com/docs/en/best-practices</a></li>
    <li id="6iQb">PASTA: <a href="https://arxiv.org/abs/2311.02262" target="_blank">https://arxiv.org/abs/2311.02262</a></li>
  </ul>
  <p id="gAIy">Follow me for more on AI and Software Development:<br /><strong>khasky</strong> — <a href="https://www.linkedin.com/in/khasky/" target="_blank">LinkedIn</a> / <a href="https://github.com/khasky" target="_blank">GitHub</a> / <a href="https://www.patreon.com/khasky" target="_blank">Patreon</a> / <a href="https://bsky.app/profile/khasky.bsky.social" target="_blank">Bluesky</a> / <a href="https://mastodon.social/@khasky" target="_blank">Mastodon</a> / <a href="https://medium.com/@khasky" target="_blank">Medium</a> / <a href="https://dev.to/khasky" target="_blank">Devto</a><br /><strong>khaskydev</strong> — <a href="https://x.com/khaskydev" target="_blank">X</a> / <a href="https://www.threads.com/@khaskydev" target="_blank">Threads</a> / <a href="https://www.instagram.com/khaskydev/" target="_blank">Instagram</a> / <a href="https://www.pinterest.com/khaskydev/" target="_blank">Pinterest</a> / <a href="https://www.tumblr.com/khaskydev" target="_blank">Tumblr</a> / <a href="https://www.facebook.com/khaskydev/" target="_blank">Facebook</a> / <a href="https://vk.ru/khaskydev" target="_blank">VK</a></p>

]]></content:encoded></item></channel></rss>