Anthropic ships official Claude Fable 5.2 prompting guide: set effort high, finish whole task
JohnAlexander · x · 2026-11-14
Anthropic published official prompting guidance for Claude Fable 5.2 and Claude Mythos 5.2 on the Claude Platform docs, covering effort calibration, early stopping, approvals, code ownership, long-turn progress updates, subagent expectations, migrating prompts designed for older models, final reports, and refusal handling.
Key takeaways:
- Effort levels: start at high (the API default), then test medium and xhigh; low is for quick tasks. Anthropic's internal tests showed medium behaves closer to Fable 5.1's higher effort.
- Stop early stopping: when long tasks pause with "should I continue?", explicitly instruct the model to finish; one user reports unfinished sessions dropped by roughly half on medium.
- Be specific in approvals: instead of replying "Go", spell out what the agent is allowed to do; "don't think" style instructions don't work at medium+, so lower effort instead.
More from coding & agent
- Lean proofs as guardrails: AI agents optimize code freely without breaking guarantees — bigdata · 2026-11-14
- jhong: biggest AI coding unlock isn't the model, it's the orchestration layer understanding user intent — gharik · 2026-11-14
- Anthropic releases free 37-minute guide to building enterprise-automating AI agents — JohnAlexander · 2026-11-14
- Snowpaw demo shows multi-agent teamwork: engineer agent fixes parser, marketing agent updates changelog — gnukeith · 2026-11-14
- Snowpaw's Scout agent shows progress driving a Fedora VM desktop — gnukeith · 2026-11-14
- Building the smallest agent with Claude Agent SDK in Python (demo) — No-Tomatillo-1938 · 2026-11-14