Two Sentences Can Make Claude Think Less. Anthropic Says When Not to Use Them. | Edition 362
Edition 362 — The fix is real, the warning matters, and six other habits can stretch your limit.
A tip is making the rounds: paste two sentences into Claude's instructions and your limit lasts longer.
The sentences are real. They come from Anthropic's own prompting guide for Opus 5.5. The promise on top of them goes further than Anthropic does.
What Anthropic found
As a chat goes on, Opus 5.5 sometimes goes back over an earlier answer while working on a new message, even a short follow-up. That adds thinking and delay. The fix tells Claude to treat answered questions as settled, starting with: “Once you have answered something, treat that answer as done.” The second sentence tells it to keep its thinking on the current question unless you ask about an earlier answer or point out a problem. Copy both from Anthropic's page.
In Anthropic's testing, follow-ups used less thinking and replies started sooner, with no effect on quality.
Notice what its write-up doesn't report: your usage limit. Effort level is one of the things Anthropic says affects your limit, so less thinking may well help. That's a reasonable guess, not a tested result. Anthropic also wrote this for developers' system prompts, so pasting it into Claude's instructions or a project is an adaptation.
The catch
Anthropic says to leave the instruction out where you want Claude re-checking its work, such as long analyses, or agent work where a later step can reveal an earlier mistake. It also warns that Claude may be less likely to point out its own earlier error.
If you use Claude for numbers, contracts, or anything you'd hate to get wrong, that's a real trade. So don't paste it everywhere. Account-wide instructions apply to every chat, while project instructions stay inside one project. Put it in a project for quick back-and-forth work, and keep it out of the one where you check figures.
Six habits with Anthropic's backing
- Lower effort first. Anthropic says this cuts thinking more reliably than prompt instructions, and medium is the default on Opus 5.5.
- Start a new chat for a new topic. Conversation length counts against your limit.
- Group related questions. Anthropic recommends putting them in one message instead of several.
- Put reused files in a project. Cached project content counts less, but the cache expires after a break, and your next message counts that content in full.
- Crop screenshots. Anthropic's API docs put a 4K image at 4,784 visual tokens and a 200x200 crop at 64.
- Use tools on purpose. Research, web search, running code, and creating files all count.
Try the screenshot math yourself:
In Claude Code: Run /clear between unrelated tasks, keep CLAUDE.md under 200 lines, and disable MCP servers you don't use. Anthropic lists all three, though it doesn't put a number on the savings.
| Claim | Figure | Source |
|---|---|---|
| Effect of the two sentences | Less thinking on follow-ups, replies start sooner, quality unaffected | Anthropic, Prompting Claude Opus 5.5 |
| Default effort on Opus 5.5 | Medium | Anthropic, Prompting Claude Opus 5.5 |
| What affects usage limits | Conversation length, effort, tools, files, model choice | Claude Help Center, Usage limit best practices |
| Image cost | 4K image 4,784 tokens; 200x200 image 64 | Anthropic, Vision docs |
| Instruction scope | Account-wide vs. one project | Claude Help Center, personalization features |
| CLAUDE.md size | Under 200 lines | Claude Code docs, Manage costs |
Try this today
Lower your effort one level on your next routine task and see if you miss anything. Then add the two sentences to one project, not your whole account. A setting you can undo beats a tip you have to take on faith.