Your Agent Does Not See the Asterisks in SKILL.md
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 🧠
The mechanism, and what the research shows
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.
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 22% on LLAMA-7B.
What the vendors document, and what I counted
Anthropic has one sentence about emphasis and it carries both halves:
If Claude keeps skipping one instruction, add emphasis such as "IMPORTANT" to that line alone. If you emphasize many lines, none of them stands out.
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.
The limits of all this
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.
The replacement
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.
Only one of those three removes the question entirely 🔧
- Anthropic's Claude Code best practices: https://code.claude.com/docs/en/best-practices
- PASTA: https://arxiv.org/abs/2311.02262
Follow me for more on AI and Software Development:
khasky — LinkedIn / GitHub / Patreon / Bluesky / Mastodon / Medium / Devto
khaskydev — X / Threads / Instagram / Pinterest / Tumblr / Facebook / VK