fable-skill
v1.0.1
Published
Claude Code skill: Fable-grade engineering discipline (explicit assumptions, scoped deliverables, evidence-backed reports) for any Claude model
Maintainers
Readme
fable-skill
A Claude Code skill that makes any Claude model (Sonnet, Opus, Haiku) exhibit the engineering discipline observed in Claude Fable 5 — explicit assumptions, scoped deliverables, evidence-backed reports.
Not roleplay. Not capability transfer. Model weights determine reasoning quality; no prompt changes that. This skill transfers the observable workflow: the output contracts and decision habits that stronger models follow natively and weaker models skip.
What it fixes (tested, not guessed)
Built with TDD for skills: baseline runs on Sonnet without the skill showed these failures; runs with the skill fixed them.
| Failure (baseline Sonnet) | With skill | |---|---| | Invents business decisions silently (stacking policies, limits, formats) | Assumptions section, one line each; asks when design-changing | | Builds unrequested scope, flags nothing | Minimum core + explicit "Extensions not built" list | | Verbose capability-prose reports | 4-part report contract, evidence first | | Success claims before running anything | "A claim with no observed output is a prediction, not a report" | | Narration preamble before the report ("Perfect! Let me...") | First-token rule: reply begins with the outcome word itself |
What it deliberately does NOT include: debug scaffolding, planning rituals, quality-gate checklists — baseline testing showed modern Sonnet already does these natively. Dead-weight instructions cost tokens and change nothing.
Install
skills.sh
npx skills add yatin-rai/fable-skillnpm
npm install -g fable-skill
# then copy the skill into your Claude Code skills directory:
# ~/.claude/skills/fable/SKILL.mdManual
Copy skills/fable/SKILL.md to ~/.claude/skills/fable/SKILL.md.
Use
Automatic — triggers on engineering tasks with ambiguity or scope risk.
Explicit: /fable <request>.
The skill detects the active model tier and scales scaffolding: Fable-class models get only the report contract (rest is native); Sonnet/Haiku get the full contracts.
Updating for future Fable generations
Prompt files cannot auto-fetch future model guidance — anyone claiming otherwise is selling magic. When a new Fable ships: run it on real tasks, diff its behavior against this spec, update the contracts. The contracts are built on enduring principles (evidence before claims, assumptions surfaced, scope bounded by the literal request), so they age slowly.
License
MIT
