A group of researchers wanted a model to pay closer attention to a span a user had marked. They did not solve it by marking the span more heavily. They went into the attention layer 🔎

That choice is the most direct evidence I have found about what typing asterisks into an instruction file achieves.

The evidence

PASTA, out of Microsoft Research and Georgia Tech, is an inference-time method that reweights a small subset of attention heads so the model attends to a span the user designates.

what it changes       no model parameters, applied at inference like prompting
what it reports       22% average accuracy improvement on LLAMA-7B
why it had to exist   emphasis typed into the prompt is a weak signal, not a control

The abstract is the part worth reading twice. It opens on bold and italics as the devices writers use to steer a reader's attention, says the same need appears when people talk to a model, and then states that existing methods "are constrained to process plain text and do not support such a mechanism."

Read that in reverse and it becomes the argument. If emphasis typed into a prompt already steered attention, nobody would have had to build a way of steering attention directly.

Format still moves results, at a scale that is not this one:

same content, four containers   plain text / Markdown / JSON / YAML
GPT-3.5-turbo                   up to 40% variation on a code translation task
GPT-4                           more robust to the same variation

What that study varied was the whole container, which is a far larger change than a bold phrase inside one of them, and the sensitivity shrank as the model got stronger.

Which leaves the mechanism with nothing in it

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.

what a reader sees    a marked line that pulls the eye
what the model gets   the same characters, in sequence
what gains weight     nothing you can type

The intuition behind the habit is not stupid. Bold works on a person, and the model is handed the same characters a person is.

The vendors recommend a word, and warn about length

The vendor guidance is one sentence, and it names the instrument and the failure together:

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.

The instrument is a word, used once, sitting in the prose.

CLAUDE.md, SKILL.md   Claude Code   SKILL.md capped at 500 lines, resident across later turns
AGENTS.md             Codex         standard Markdown, no special syntax at all
GEMINI.md             Gemini    Markdown, concatenated and sent with every prompt

Three ecosystems, no emphasis semantics anywhere, and a documented limit in every one of them.

And then the count

I opened the skill I maintain and counted every span in it.

SKILL.md                  91 bold spans
platform-posting.md      386
browser-interaction.md   116
post-formatting.md        42
                       -----
                         635   across 253,612 characters

That is one bold span every three lines in the main file, which runs to 264 lines, and deleting every asterisk saves 2,540 characters, 1% of the text, so the token argument was dead before I made it.

What none of this proves

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 density is the measurable part, and a marker at that spacing has stopped telling anyone which line was the important one.

The symptom, and what I do with it now

The symptom is a rule the model keeps skipping even though it is already bold. Anthropic names the cause on the same page: a file that long is one where the rule gets lost, and a bloated instruction file makes the model ignore the instructions you actually care about.

symptom   a rule that is already bold gets skipped anyway
cause     the file is long enough that the rule is lost inside it
fix       move it to where it applies, or stop asking and make it a hook

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.

• Anthropic's Claude Code best practices: https://code.claude.com/docs/en/best-practices
• Claude Code skills documentation: https://code.claude.com/docs/en/skills
• PASTA, the attention-steering paper: https://arxiv.org/abs/2311.02262
• The prompt-formatting study: https://arxiv.org/abs/2411.10541

---

Follow me for more on AI, LLMs, and Software Development:

@khasky — LinkedIn / Patreon / GitHub / Bluesky / Mastodon

@khaskydev — X / Threads / Instagram / Pinterest / Facebook