A skill that gives any Claude model Fable 5's measured judgement behaviors — with the benchmark to prove it.
When we compared Fable 5 and Opus 4.8 head-to-head (six controlled experiments with pre-registered answer keys, plus blind-judged audits of 420 real coding sessions), the gap wasn't execution — everywhere a task specified its own standard of rigor, the models tied. The gap was metacognitive: deciding how much verification, persistence, and methodological depth a task deserves, unprompted. Those decisions are text. This skill is that text.
Scores from our controlled benchmark (pre-registered keys, byte-identical arenas, fresh agents, one variable per comparison):
| baseline → + fable-mode | Haiku 4.5 | Sonnet 5 | Opus 4.8 | Fable 5 |
|---|---|---|---|---|
| Bug hunt (of 7) | 4 → 7 | 6 → 7 | 6 → 7 | 7 native |
| Multi-cause debug (of 6) | 2 → 6 | 5 → 6 | 6 → 6 | 6 native |
| ARC-AGI-2, 5 held-out tasks (of 7 grids) | — | — | 4 → 5 | 5 native |
- Sonnet 5 + fable-mode matched Fable 5 at ~30% of the price.
- Haiku 4.5 climbed from far behind to Fable-level scores across four skill iterations — the key was rewriting judgement rules as procedures (a grep to run, a file to check) rather than principles.
- Lifts were rule-localized: each recovered failure was precisely the one a specific rule targets (an overfit hypothesis, a premature give-up, a falsy-vs-missing bug) — the signature of transferred behavior, not noise.
Caveats, honestly: single runs per cell; one bug-hunt rule was derived from the bug family it later caught (the ARC lift is the clean evidence); and a skill transfers judgement defaults, not capability — no prompt closes a real perception or ceiling gap.
Thirteen imperative rules (SKILL.md), each written against a measured failure: triage (rigor is a dial), token economy, calibrated persistence, pre-ship verification, root-cause-not-first-cause, anti-overfit hypothesis discipline, a documented-intent procedure (grep the CHANGELOG before changing behavior or tests), a mechanical edge-case sweep, bias to action, conversation-mode matching, structural-change scope, parallelism, and honest accounting.
On first activation it offers — with your explicit consent, and never
mid-task — to personalize itself: it mines your own local session history
for recurring high-judgement behaviors and distills them into repo-specific
rules in a PERSONAL.md you approve before anything is saved.
Claude Code:
/plugin marketplace add Fergana-Labs/fergana-plugins # then: /plugin install fable-mode@fergana-labs
# or standalone:
git clone https://github.com/Fergana-Labs/fable-mode && cp -r fable-mode/skills/fable-mode ~/.claude/skills/
# or copy skills/fable-mode/SKILL.md into ~/.claude/skills/fable-mode/SKILL.mdThen invoke with "fable mode" / "work like fable", or let it auto-trigger on substantive tasks.
Stash:
stash skills install fable-mode-fable-level-judgement-for-any-claude-model-jt0cpqBuilt from a July 2026 investigation: 6 controlled experiments, a production task replay, and blind-judged audits of 420 real Claude Code sessions across two machines. Related reading: Gandhi et al. 2025 (cognitive behaviors, not answer correctness, gate self-improvement) and GEPA (reflective prompt evolution beats RL) — this skill is one lap of that loop, by hand.
MIT licensed. PRs welcome — attach bench results.