# What to Stop Prompting Now

*Anthropic's own guidance for its Claude 5 models says to delete lines most of us still type: think carefully, show your reasoning, retype your charts. Here is the stop-doing list, and what to write instead.*

addAI.dev · Field Notes

By [Jay](mailto:jay@addAI.dev) · October 2026

`#AITools` `#TechDeepDive` `#HotTake`

Most of the prompting habits people carry around were **correct advice** for the models they were learned on. Tell the model to think step by step, to show its work, and the answers got better. Anthropic has now published two guides, about two months apart, saying that for its newest models a good share of those habits should be deleted.

The first is Thariq Shihipar's [The new rules of context engineering for Claude 5 generation models](https://claude.dev/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models/) (July 24, 2026). The second is Addy Osmani's [Getting the most out of Opus 5.5 in Claude and Claude Code](https://claude.dev/blog/getting-the-most-out-of-opus-5-5/) (September 22, 2026). Together they make a stop-doing list, and each item on it was good advice until recently. Lines like these sit in saved instructions, custom GPTs and assistant setups at businesses of every size, quietly adding cost and removing quality.

> We removed over 80% of Claude Code's system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations.

-- Thariq Shihipar, Anthropic, July 24, 2026

## Stop telling it to think

Anthropic's instruction is to remove "think carefully," "think step by step," and similar lines from your prompts and your saved instructions. The reason given is that Opus 5.5 always thinks before it replies and decides how much on its own. Anthropic's test is stated narrowly, so it is worth quoting narrowly:

> In our testing in a chat product, removing a “think carefully” line made replies start sooner, with no clear drop in quality.

-- Addy Osmani, Anthropic, September 22, 2026

The fix is not more prose. For a simple question you can say *Answer directly*, and in Claude Code you change the **effort** setting, the control for how much the model thinks. Anthropic's own [prompt-audit catalogue](https://github.com/anthropics/skills/blob/main/skills/claude-api/shared/prompt-audit.md) makes the same point from the API side: where thinking is always on, effort is the only thinking control, and it cuts thinking, cost and latency more reliably than prose does. It adds that a reply-directly line is worth keeping only where a measurement on a latency-sensitive route shows it helps.

## Stop asking it to show its reasoning

This one carries a real cost. Anthropic's guide says a request to reproduce the model's internal reasoning in the reply **can be declined**, and that it is one of the flag categories. The audit catalogue says the same for the API on Opus 5.5, Sonnet 5.5 and Fable 5.1: instructing reasoning reproduction can trigger a refusal. What to ask for instead is the thing you actually want, for example *Explain why you chose this approach in three sentences*, which is the guide's own example.

## Stop repeating yourself in chat

If you catch yourself typing the same correction for the third time, the chat is the wrong place for it. A standing rule belongs in the file the tool reads at the start of every session, which in Claude Code is `CLAUDE.md`. Anthropic's guide puts a keep-going and stop-and-ask rule there.

Long runs need the same treatment for progress. The guide has Claude keep a checklist in `TASKS.md`, because a long run fills the context window (the session's working memory) and Claude Code then summarizes older turns. In the guide's words, "A list in a file survives that."

Which file you use matters, and we tested it. In [Claude Code Imports AGENTS.md - Until It Doesn't](/blog/claude-code-agents-md.html) we ran `/compact`, the command that summarizes a long session, on our own setup and checked what came back. `CLAUDE.md` itself was re-read, and the content behind an `@AGENTS.md` import inside it **did not come back** in our test. Standing rules belong **in the instruction file itself**, not in the chat and not one import away.

## Stop asking for an outline

If you need a spreadsheet or a document, ask for the spreadsheet or the document. Anthropic's reason is that what Opus 5.5 produces needs less editing before you share it than Opus 5's did, so the outline-first step now adds a round trip and nothing else.

## Stop retyping charts

Attach the chart, diagram or screenshot and ask a specific question about it. Anthropic's audit catalogue lists chart-reading step lists and OCR pre-passes as scaffolding for weaker vision, and says Opus 5.5 reads charts, diagrams and screenshots considerably more precisely without them.

## What to write instead

The replacement for most of that list is one move: say **what done looks like**, in a single message, and let the model run. Anthropic's example is a migration task that ends with "Done means: every endpoint uses the new client, the old client is deleted, and the test suite passes," followed by when to stop and ask. The stop-and-ask rule from the section above comes with one safeguard: permission prompts stay on for anything destructive.

The same guide applies the idea to design requests. *Avoid a generic look* mostly swaps one default for another, so the guide's example names the patterns to exclude: a cream background, italic accent words in headings, numbered section labels, monospace labels, pill-shaped buttons. A vague negative is out. A specific one is in.

## Test the removals

Two limits apply. First, Anthropic's "no clear drop in quality" came from a chat product, not from every task. Second, the guidance is about Claude 5 family models, and older models or other vendors' models may still respond to the old habits.

> test removals, not just additions

-- Anthropic's prompt-audit guide, on patch accretion

That fragment, from the audit guide's fix for patch accretion, is the method. A prompt that has collected every generation's workarounds is hard to reason about, and a better score on your test cases from adding one more line on top is not evidence that the stack should exist. Delete a line, re-run the task that line was written for, and keep it only if the failure comes back. We ran Anthropic's audit across seven of our own projects in [When Your AI Agent's Instructions Contradict Themselves](/blog/when-instructions-contradict-themselves.html).

## What this means for your business

If your team has been writing prompts since the GPT-4 era, your saved instructions, custom GPTs and agent configuration files are carrying lines like these right now. Each one costs tokens on every request, and some of them make the answers worse. addAI.dev audits those files against the current guidance, removes what the model no longer needs, and re-tests every change before it ships, so the AI you already pay for answers faster, costs less per request and holds its quality.

[Get in touch](/contact.html)