Anthropic's Claude Fable 5.1 prompting guide: 15 behavior changes with fix prompts

AGI Hunt · wechat · 2026-09-02

Alongside Fable 5.1, Anthropic published an official prompting guide, noting many old prompts merely patched older models' bad habits, and listing behavior changes with copy-paste fix prompts.

Key behavior changes: fewer parallel tool calls (may degrade to one tool per turn), fewer progress updates during long calls, answering from memory under low effort, denser prose, less formatting in chat, summaries without quotes, and full-file rewrites for small edits. Each has an official fix, e.g. "privately list what you need next, then request all independent items in one response" to restore parallelism, and "you are operating autonomously" to stop mid-task questions.

Effort tiers (low/medium/high/xhigh/max, default high): start at high and tune down via your own evals; on CursorBench 3.2.0, 5.1's low matches 5's high at 1/3 the cost; cache reads drop to $0.25/MTok and effort can be switched mid-conversation without breaking cache.

Context management: treat history as append-only — editing earlier turns invalidates thinking blocks (enforced for new accounts after Aug 31); with cheaper cache reads, client-side early compression may no longer pay off. The guide also defines "mannered prose" with a removal prompt, a six-item compaction checklist (keep blockers/solutions, rejected options, decisions, progress, todos, specifics; preserve user wording verbatim), and a constraint prompt limiting unrequested code changes.

Original post →

Anthropic's Claude Fable 5.1 prompting guide: 15 behavior changes with fix prompts

More from coding & agent