fable-skills

A skill library that teaches Claude Opus the working disciplines of Claude Fable 5.

Drop these skills into any Claude Code setup running Opus (or Sonnet, or any earlier Claude model) and the model picks up the habits that make Fable feel different: it verifies before it claims, it finds the root cause before it fixes, it leads with the outcome when it writes, and it finishes the turn instead of ending on a promise.

Unofficial. This project is not affiliated with or endorsed by Anthropic. It is a community distillation of publicly documented model behavior.

The thesis

A skill file cannot give a model more capacity. Fable's raw advantages, like first-shot correctness on complex problems and instruction retention across multi-day runs, do not transfer through markdown.

But a surprisingly large share of the observable difference between Fable and Opus is process, not capacity: what the model checks before it speaks, when it stops working, how it decides something is done, and what it cuts from a draft. Those are encodable. Anthropic's own Prompting Claude Fable 5 guide documents these behaviors as short instruction blocks, and notes that skills written for older models are often too prescriptive for Fable. The inverse is the whole idea of this library: older models benefit from exactly the prescriptive scaffolding Fable no longer needs.

One line summarizes every skill here: you cannot give a model more capacity with a markdown file, but you can move its checkpoints. Fable pauses and audits before producing output. Opus produces output first. These skills move the checkpoint.

The skills

Skill	Domain	The discipline it enforces
verified-done	Software	Never claim completion without tool-result evidence. Report failures faithfully.
root-cause-first	Software	Reproduce and trace before fixing. Never retry a failed approach verbatim.
minimal-diff	Software	Do the simplest thing that works well. No unrequested refactors or speculative abstraction.
delegate-and-verify	Software / agents	Fan out independent work to subagents. Verify with fresh-context checkers, not self-critique.
finish-the-turn	Agent behavior	Act when you have enough information. Never end a turn on a promise.
lessons-ledger	Agent behavior	Record one lesson per file across runs. Update, prune, and actually reference them.
outcome-first-writing	Writing	Lead with the outcome. Shorten by selectivity, never by compression.
plain-handoff	Writing / agents	Write final summaries for a reader who did not watch the work. Retire invented shorthand.
evidence-audited-analysis	Data science	Interrogate the data before analyzing it. Reproduce headline numbers. State what the data cannot show.

Install

Claude Code, all projects (user level):

git clone https://github.com/benjaminard/fable-skills.git
cp -r fable-skills/skills/* ~/.claude/skills/

Claude Code, one project:

cp -r fable-skills/skills/* your-project/.claude/skills/

Each skill is a plain-markdown SKILL.md with standard frontmatter, so the same files work anywhere Agent Skills are supported, including the Claude Agent SDK.

Install the whole set or cherry-pick. The skills are independent by design, though verified-done and finish-the-turn are the two with the highest payoff per token if you only take two.

What to expect

These skills reliably change behavior that is process-shaped:

• Fabricated or optimistic status reports drop sharply. Anthropic reports that the claim-auditing instruction these skills encode "nearly eliminated fabricated status reports" in their testing on Fable; the same discipline transfers.
• Symptom-patching gives way to root-cause fixes, and repeated identical retry loops stop.
• Diffs shrink. Unrequested refactors, defensive error handling, and drive-by cleanup mostly disappear.
• Written output leads with the answer and reads like prose instead of fragment stacks and arrow chains.
• Analyses start with data profiling and end with explicit uncertainty instead of confident narrative over unexamined numbers.

They will not close the gap on raw reasoning depth, long-horizon coherence, or first-shot correctness. A hard problem that Opus cannot solve remains a hard problem with these skills installed. What changes is that Opus will tell you it has not solved it, instead of telling you it has.

Contributing

Additions welcome if they meet the bar: the skill must encode a checkpoint (something the model verifies or decides before producing output), not a preference. One discipline per skill. Keep each SKILL.md under roughly 100 lines; if it needs more, it is two skills.

License

MIT