Anthropic's Fable 5.2 prompting guide: effort, approvals, code ownership

Anthropic published a prompting guide for Claude Fable 5.2 and Claude Mythos 5.2 on its platform documentation site, and Simon Willison quoted its code-ownership section on 10 November 2026. The guide follows the 9 November release of both models. It is vendor documentation: every figure comes from Anthropic's own testing, and it does not publish the task sets behind them.

Effort: start at high, rerun your sweep

According to the guide, high is the default on the Claude API and most of the capability gain is already present there; the step from medium to high raised scores on most of the coding and knowledge-work evaluations Anthropic tracks, at a cost in tokens and time. Anthropic suggests low for high-volume work and subagents, medium for interactive coding and chat, and xhigh only where you have measured a gain.

It also warns that effort levels do not map one-to-one across models, so a setting tuned on Fable 5.1 or Opus 5.5 should be re-tested. At high and above the model thinks before almost every reply, including a greeting, and in Anthropic's testing an instruction to answer without deliberating did not stop that at medium and above.

Early stops and approvals

At low and medium effort the model sometimes ends a long task early, with a list of what remains or an offer to continue. Anthropic supplies a 'Finishing work' system prompt block. On recorded sessions from an agentic coding product at medium effort, a version of it roughly halved the sessions that ended with doable work left, and raised output tokens by about 30 percent. The guide notes it removed few endings that list what remains and then stop.

On approvals, the guide says to reply with the exact action, such as 'Yes, run the migration on the staging database', not 'go'. If your system prompt has approval rules, name the actions they cover: in one simulated test a rule naming messages to people outside the conversation made the model hold back about half of its routine requests to colleagues.

Code ownership: a trade, not a free gain

Anthropic says Fable 5.2 follows what the prompt says about who owns the code. With the sentence 'This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it', the model merged repeated code in about half of tries on small coding tasks, against under a tenth without it.

The guide lists the cost. Output tokens rose about 1.5 times, the model added new tests and documentation nobody asked for, and in one try in six to ten it changed how an existing feature behaves. Its closing summary named the larger changes but often omitted small ones, so Anthropic advises reviewing the diff, not the summary. A 'no drive-by refactors' rule elsewhere in the prompt did not cancel the sentence; merging still happened in about a third of tries.

Cleaning up inherited prompts

For teams migrating, the guide names lines to reconsider: standing instructions to delegate to subagents, which can drive up session cost; token budgets written into the prompt, where removing a large one lowered cost with equivalent performance at high and xhigh effort; and 'think less' lines. It adds that notes and memory files accumulate cautions, and in one simulated setup stored 'never retry' notes cut retries from nearly all tries to about four in ten.

One operational constraint: Anthropic says the API now enforces a prefix check for every account on Fable 5.2, so changing the system prompt, tools or an earlier message mid-conversation can return a 400 error when thinking blocks are sent back. Apply prompt changes in new conversations, or put new instructions in the newest turn.

What the guide leaves out

The guide gives rates but not sample sizes, task lists or confidence intervals, and the tests are on small coding tasks and simulated workflows. Whether the same sentences behave the same on your codebase is for your own evals to answer.