Fable Brain

Fable 5 Best Practices Cheat Sheet

5 min read

Fable 5 Best Practices Cheat Sheet

The fifteen highest-leverage verified practices for operating Claude Fable 5 across model choice, prompting, context, and safety, each with a source and a confidence tag.

What it is

  • A one-page operator cheat sheet distilled from the vault’s verified claim packs of 2026-07-07.
  • Scope: the model itself (choice, effort, refusals, cost) plus the two main harness disciplines, prompting and Claude Code operation.
  • Every practice below is a single claim with a single tag; deeper treatment lives in the linked notes.

How it works

  • Claims were verified against dated official pages (Anthropic news, Claude Platform docs, code.claude.com docs) and tagged per Confidence Tag Policy.
  • EVIDENCE-BASED means a dated public Anthropic URL corroborates the practice; PRACTITIONER means converging field practice without official confirmation.
  • Sources are cited inline per practice and collected under Sources.

Best practice

  1. Default to Fable 5 when capability dominates cost; Anthropic states “Fable 5’s capabilities exceed those of any model we’ve ever made generally available.” (https://www.anthropic.com/news/claude-fable-5-mythos-5). EVIDENCE-BASED
  2. Control depth with the effort parameter, not thinking budgets: high is the default, xhigh serves the most capability-sensitive workloads, and budget_tokens returns a 400 error on Fable 5 (https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking). EVIDENCE-BASED
  3. Lower effort before switching models for cost control; “Lower effort settings on Claude Fable 5 still perform well” and often exceed xhigh performance of prior models (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5). EVIDENCE-BASED
  4. Architect for refusals: handle stop_reason “refusal” with its stop_details category, and enable Claude Opus 4.8 fallback via the server-side fallbacks parameter (beta header server-side-fallback-2026-06-01) or SDK middleware; a refusal before any output is unbilled (https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback). EVIDENCE-BASED
  5. Exploit the 90% prompt-caching input discount against the 50 per MTok list price on repeated-context workloads (https://www.anthropic.com/claude/fable). EVIDENCE-BASED
  6. Write clear, explicit instructions with numbered steps when order matters, and request above-and-beyond behavior explicitly (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview). EVIDENCE-BASED
  7. Give the motivation behind an instruction; Fable 5 performs better when told why a request matters, since “Claude is smart enough to generalize from the explanation.” (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5). EVIDENCE-BASED
  8. Structure complex prompts with consistent XML tags and wrap 3 to 5 relevant, diverse few-shot examples in example tags (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview). EVIDENCE-BASED
  9. In long-context prompts (20k+ tokens), put documents at the top, the query at the end, and ask for grounding quotes first; end-placed queries improved response quality “up to 30% in tests” (https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview). EVIDENCE-BASED
  10. Never instruct Fable 5 to echo or transcribe its internal reasoning in response text; that can trigger the reasoning_extraction refusal category. Read structured thinking blocks instead (https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking). EVIDENCE-BASED
  11. Refactor prompts and skills written for earlier models rather than copying them; overly prescriptive instructions can degrade output quality on the stronger model (https://knightli.com/en/2026/06/10/claude-fable-5-prompting-guide/). PRACTITIONER
  12. In Claude Code, give the model a verifiable check so the loop closes without you: “Give Claude a check it can run: tests, a build, a screenshot to compare.” (https://code.claude.com/docs/en/best-practices). EVIDENCE-BASED
  13. Follow explore, plan, implement, commit, but skip the overhead for small fixes: “If you could describe the diff in one sentence, skip the plan.” (https://code.claude.com/docs/en/best-practices). EVIDENCE-BASED
  14. Keep CLAUDE.md short and broadly applicable, target under 200 lines per file, and move procedures into skills that load on demand (https://code.claude.com/docs/en/memory). EVIDENCE-BASED
  15. Practice context economy: /clear between unrelated tasks, delegate scoped research to subagents, and after roughly two failed corrections start a fresh session with a better prompt (https://code.claude.com/docs/en/best-practices). EVIDENCE-BASED

Pitfalls

  • Enforcing mandatory rules with prose instead of hooks; hooks are deterministic, and policy hooks must exit 2 because exit 1 does not block (practitioner consensus, https://blakecrosley.com/blog/claude-code-hooks-explained).
  • Forgetting Fable 5 is a Covered Model: 30-day data retention for safety monitoring is mandatory and zero data retention is unavailable, which matters for compliance-sensitive deployments.
  • Budgeting from anecdote: day-one testing found Fable 5 slow and expensive ($110.42 in about 5.5 hours), but that figure is a contested single-practitioner data point.
  • Assuming safety classifiers rarely matter; Anthropic reports they trigger in under 5% of sessions on average, which is still frequent at production volume.
  • Aggressive CRITICAL/MUST phrasing carried over from older prompts causes tool overtriggering on current models.

Sources

Next actions

  • Re-verify all fifteen practices at the next brain refresh against their source URLs.
  • Add a PRACTITIONER section if field consensus diverges from official guidance.
  • Test practice 3 (effort downshift) on a real workload and record results.

Graph View

Backlinks