v2.0.0 — Make your model behave like Fable
From postures to gates. v2 rebuilds fable-it from firsthand Fable 5 research (live harness introspection + Anthropic's "Prompting Claude Fable 5" guide), shipped against a 26-case binding test contract — tabletop goldens judged by fresh-context agents, scripted lints, hook unit tests, and live [REAL] runs on Sonnet 5 and Opus 4.8.
Highlights
- 5-gate catalog replaces posture prose: turn-end, claim, state-change, phase-boundary, delegation — each a trigger + test + action
- Evidence ledger —
VERIFIEDis a lookup against.taskstate/evidence.md, not a vibe (the Anthropic-measured claim-grounding rule) - Fresh-context verifier audits every report before it ships (reads only the DoD, the report, and the ledger)
- Disk-backed run state that survives compaction, re-read at every phase boundary + cross-run lessons
- Model-adaptive posture for Sonnet 5 and Opus 4.8, applied at Step 0
- Cost-aware delegation routing — packets routed to model tiers by task shape, spend disclosed per agent
- /launch unattended mode — overnight runs never stall on an approval prompt (human runs keep their gates)
- Hardened mode (opt-in, fail-open) — hooks that mechanically block promise-endings and evidence-free VERIFIED rows
- CDP core single-sourced + route guard — no autonomous write path on authenticated sessions, ever
- Portable root SKILL.md — full mechanics (not descriptions) for Cursor, Codex, Copilot and 70+ agents
Full details: CHANGELOG.md · per-epic board: delivery/STATUS.md · research: docs/research
Install / upgrade
/plugin marketplace add DevOtts/fable-it
/plugin install fable-it@devottsOther agents: npx skills add DevOtts/fable-it -a <agent>
🤖 Generated with Claude Code