Deep dive · XiaoHu explains

Anthropic's official Opus 5 prompting guide is mostly about what to delete

Eight prompts you can copy straight into production, plus one counterintuitive catch: tell it to report only high-severity issues and it may genuinely report fewer.
The one-minute version
  • Anthropic published a page devoted to one question: which lines of your existing prompts need to change once you move to Opus 5. The strangest part is that three of its eight sections are about taking things out
  • Top of the delete list: "add a verification step at the end" and "use a subagent to double-check." Opus 5 already verifies itself, so these carryovers from the previous generation just make it verify twice — more tokens, same quality
  • Code review prompts need the opposite instinct. Write "report only high-severity issues" or "be conservative" and it may take you literally. The fix is to let it report everything and filter in a separate pass
  • It talks more than its predecessor, narrates more, and delegates to subagents more eagerly. The doc supplies eight copy-paste prompts, one per symptom
  • Thinking is now on by default. You can only disable it at high effort or below (low through max are the five effort levels; full table in section 8) — pair disabled thinking with xhigh or max and the API returns a 400
⚑ The material comes from Anthropic's own documentation, describing how to tune its own model — capability claims and recommended settings are self-reported. The five-level breakdown, error details, and cache minimum come from the Effort docs and migration guide on the same site; the text flags each one where it appears.
1 What this is

A document that tells you to delete things

If your system prompt contains a line like "after completing the task, add a verification step," it started costing you money the moment you switched to Claude Opus 5.

Anthropic wrote a standalone prompting page for Opus 5 — eight sections, a bit over 1,500 words. Three of those sections are about removing things from your prompt.

The same system prompt, after moving to Opus 5
−
Add a verification step at the end of any non-trivial task.
Use a subagent to double-check the result.
Re-check your answer before responding.
+
Keep responses as long as they need to be, no longer.
Delegate only work that's genuinely large and parallelizable.
Deliver what was asked; don't expand the scope yourself.
All three lines on the left were written for the previous generation. Opus 5 folds those behaviors into its defaults, so keeping the lines just makes it do the work twice.

The reversal isn't hard to follow. For years the default move in prompt engineering was addition: the model wouldn't check itself, so you wrote a line forcing it to; it delivered half-finished work, so you wrote a line forcing it to finish; it wouldn't verify, so you hung a subagent off the end to review. Every one of those lines patched a weakness.

Opus 5 closed a batch of them. It verifies on its own, finishes the whole task instead of handing back a stub (no placeholder functions with a name and nothing inside), and goes back to fix its own mistakes when it spots them. Those forcing lines now force behavior it would produce anyway.

These instructions cause Claude Opus 5 to over-verify. Removing them reduces wasted tokens with no loss in quality.Prompting Claude Opus 5 · Anthropic

Worth separating this page from the Opus 5 launch announcement — they're different documents. The announcement covers how capable the model is, what it costs, and whether it's worth switching. This page covers exactly one thing: once you've switched, which lines in the prompts you inherited from the Opus 4.7 and 4.8 era need editing.

Related on this site
Claude Opus 5: half the price, close to Anthropic's own top model Fable 5 — but still behind on 5 of 12 benchmarks
2 Delete

Start by cutting every line that tells it to check its own work

This is the first thing to do in the whole document. The lines you're hunting for come in two flavors, and both will look familiar.

The first is task-level verification: "include a final verification step for any non-trivial task," "use a subagent to verify." The second is response-level self-checking: "double-check your answer," "confirm before you respond." Same treatment for both — delete.

One instruction, two generations of model
Previous generation: without this line, it really doesn't verify
Do the work
→
The pass you asked for
→
Deliver
Here the line earns its keep — it supplies the step the model was missing.
Opus 5: it already verified
Do the work
→
Its own pass
→
The pass you asked for
→
Deliver
Same output, one extra lap. The cost of that lap is real; the return is zero.

The cleanup goes beyond your prompt text. The doc calls this out specifically: if the harness you built for older models contains a separate verification stage, tear that out too. The prompt is what you say to the model; the harness is the code wrapped around it that decides when to call the model, how many times, and with what. Both layers need clearing.

Where this lands for Claude Code users

The page is written entirely from an API perspective — system prompts and outer harnesses. If you're using Opus 5 inside Claude Code, the equivalent surfaces are CLAUDE.md and your own skill files. The lines to look for read like this: "after editing code, have a subagent review it," "verify each step before moving on," "reconfirm all changes before delivering." All three should go. The doc doesn't spell out this layer; it's our extrapolation, same model and same system prompt.

3 Invert it

Ask a review prompt for "only the serious ones" and you'll get fewer

This is the least intuitive item in the guide, and the easiest one to trip over inside your own automation.