fable-5

Claude Fable 5 (claude-fable-5) protocol — selection, prompting, effort tuning, token discipline. Trigger when working with, routing to, or writing code that calls Fable 5.

SKILL.md

GitHub

Fable 5 — Maximum-Capability Protocol

Previous generation. Fable 5.1 (fable-5-1) succeeds it at the same price with breaking API changes (forced tool use rejected, preserved thinking) — prefer that skill for new work; use this one for projects pinned to claude-fable-5.

Model ID: claude-fable-5 · 1M context (default and max) · 128K max output · $10 in / $50 out per MTok — the most expensive tier. Every rule below exists to convert that cost into results a cheaper model can't produce.

Staleness guard

Model facts here (ID, pricing, limits, API rules) were verified 2026-07. If a newer Claude model generation exists, pricing looks off, or the API contradicts this file: verify against current docs (Models API / platform.claude.com or the claude-api skill) before relying on them, follow the live source, and log the correction in learnings.md.

Self-improvement (do this first and last)

  1. At start: read learnings.md in this skill's folder if it exists. Apply every relevant lesson.
  2. At end of every use: append one dated bullet to learnings.md — what worked, what wasted tokens, what you'd prompt differently. One lesson per bullet, concrete, project-agnostic. Merge into an existing bullet instead of duplicating; delete bullets proven wrong.

When to route work here

Use Fable 5 only for work at the top of the difficulty range: long-horizon autonomous runs, hard architecture/debugging that other models failed, first-shot builds of well-specified systems, end-to-end deliverables. Routine work belongs on Opus 4.8 or below — Fable at low effort still beats prior models' max, but you're paying 2× Opus rates for it.

Hard API rules (violations are 400 errors)

  • Omit the thinking parameter entirely (or {type: "adaptive"}). Thinking is always on; disabled and budget_tokens are rejected.
  • No temperature / top_p / top_k. No assistant prefill.
  • Raw chain of thought is never returned. thinking: {type: "adaptive", display: "summarized"} for readable summaries; default is omitted (empty thinking text).
  • Handle stop_reason: "refusal" before reading content, and ship the server-side fallback by default: betas: ["server-side-fallback-2026-06-01"], fallbacks: [{"model": "claude-opus-4-8"}].
  • Requires 30-day data retention (ZDR orgs get 400 on every request).

Getting the most out of it

  • Full spec up front, one turn. State the goal, constraints, and what "done" looks like in the first message. Progressive drip-feeding reduces quality and burns tokens.
  • Give the reason, not just the request: "I'm working on X for Y; they need Z. With that in mind: [request]."
  • Effort is the lever: output_config.effort — high default, xhigh for the hardest coding/agentic work, max only when correctness outweighs cost, low/medium for routine passes. Lower effort = fewer, more consolidated tool calls.
  • De-prescribe. Prompts written for older models (step-by-step scaffolding, "CRITICAL: YOU MUST") reduce Fable output quality. State goal + constraints, not steps.
  • Delegate. Parallel sub-agents are dependable — fan independent subtasks out and keep working; asynchronous beats spawn-and-block.
  • Give it a memory surface. Even a plain .md file it can write learnings to measurably improves multi-session work.
  • Plan for long turns. Single hard requests can run many minutes — stream, set generous timeouts, don't block a UI on one call.

Token discipline

  • Anti-overengineering guard: "Don't add features, refactor, or abstract beyond what the task requires. Simplest thing that works."
  • Ground progress claims: "Only report work you can point to tool-result evidence for."
  • For autonomous runs, add the no-early-stop reminder: act instead of announcing intent; end the turn only when done or blocked on the user.
  • Cache-friendly prompts: frozen system prompt first, volatile content after the last cache breakpoint.