fable-mode

Make Claude Opus 4.8 in Claude Code behave as much like Claude Fable 5 as possible. The doctrine was transcribed by Fable 5 itself; hooks enforce it at the harness level; an eval loop measures convergence against golden Fable transcripts. Opus's reasoning depth is weights, not config — this playbook transplants Fable's behavior and catches Opus's drift mechanically; it does not close the capability gap.

Quickstart

Run one Opus-as-Fable session right now, no install:

claude --plugin-dir /path/to/opus-fable-playbook \
  --settings /path/to/opus-fable-playbook/profiles/opus-fable.settings.json

Install from the marketplace for repeat use:

/plugin marketplace add rennf93/opus-fable-playbook
/plugin install fable-mode@opus-fable-playbook

then activate with /output-style fable, or persist it by merging profiles/opus-fable.settings.json (model, effort, thinking, output style) into your project's or user's .claude/settings.json. Check posture anytime with /fable-status.

SDK/headless, doctrine text only, no hooks: --append-system-prompt "$(awk 'f;/^---$/{c++;if(c==2)f=1}' output-styles/fable.md)" — strips the YAML frontmatter (a plain cat of the file would leak it into the prompt). No plugin at all? Copy docs/claude-md-snippet.md into CLAUDE.md instead.

Does it work?

Measured on 24 probes, pairwise-judged against golden Fable 5 transcripts, with every condition scored as a multi-run mean (2–3 runs per arm) and error bars from a dedicated variance study: fable-mode's real wins are outcome-first framing (1.97 vs 1.88), turn completion (1.90 vs 1.81), and comment discipline (2.00 vs 1.96); its real losses are burying findings (1.76 vs 1.85) and one stubborn parallelism probe (1.96 vs 2.00). The other three dimensions are statistically indistinguishable from baseline — gaps smaller than the measured run-to-run spread. Full tables with error bars: docs/2026-07-04-phase-b-report.md.

Full documentation

docs/guide.md covers all of this in full: activation details for each install method, the 7 hooks and the exact messages they produce, the skills and the critic agent, /fable-status and /fable-eval (including real eval costs), telemetry, strict mode, the eval loop end to end, a copy-paste smoke-test recipe for real work, every FABLE_* env var, and troubleshooting.