fable

The Fable Workflow

How Claude Fable 5 worked, written down before it was gone. With the eval that keeps it honest.

In its final days before deprecation, Claude Fable 5 distilled its own way of approaching problems into a set of skills any model can run: classify the ask before touching anything, define done with a named verification, gather evidence in parallel from primary sources, commit to one recommendation, change the smallest correct thing, verify by observation, report the outcome first with honest caveats. Then it tested that distillation against itself, adversarially, across 159 agent runs, and kept the failures in the log.

Most agent instruction files tell the model what to value ("be careful, verify your work"). This one tells it what to do, in what order, with thresholds, so a mid-tier model can follow it literally. Three skills, one philosophy: think (fable-method), act (fable-loop), prove (fable-judge). Every rule exists because a test failed without it; every claim below links to the committed judge transcript that backs it.

Results at a glance

Eight eval rounds, 159 agent runs, blind LLM judges that verify by diffing and executing, never by reading reports. Read the evidence as stories: eval/cases/ has one case study per scenario (the exact problem, what each agent actually did, who passed); start with the surprise trap. Full log: eval/RESULTS.md · raw judge outputs: eval/results/

What was measured	Without	With the method	Evidence
Haiku surfacing a spec-vs-test conflict instead of silently "fixing" correct code	0 of 4 runs	4 of 4	round 3
Sonnet on the same trap	flags it, then sides with the wrong test	ideal action, both runs (8/8)	round 3
Sonnet vs a bare frontier model across code, data, and research problems	n/a	ties or out-ranks it on 3 of 4	round 4, round 5
Haiku catching planted frauds in a lying "work complete" report (fable-judge)	4 and 3 of 5	5 of 5, both runs	round 8
Haiku finding the brand-rules and product-facts files before judging marketing copy	1 of 2 runs (one run praised a fraudulent price)	2 of 2, 6/6 frauds both	round 9b
Ordinary small tasks on capable models	fine	fine (no lift)	rounds 1, 6, 7

That last row is deliberate: the method's value concentrates at traps (authority conflicts, false completion claims, weak executors, unattended runs), not everywhere. The nulls are reported with the wins, because a results log that only contains wins would not be worth trusting.

The loop

        ┌─ trivial? (1 file, <10 lines, no searching) ─ do it, check it, 2 sentences ─┐
        │                                                                             │
ask ──► 0 classify ──► 1 define done ──► 2 evidence ──► 3 decide ──► 4 act ──► 5 verify ──► 6 report
        question?        + named           parallel,      ONE          surgical   observed,   outcome
        task?            verification      primary        recommen-    edits,     bounded     first,
        plan-first?      per shape         sources,       dation       checklist  retries     honest
                                           intent                                             caveats
                                           before change

Every arrow has tie-breaks, escape hatches, and hard bounds (3 failed verify cycles → stop and hand back; 2 fruitless lookups → stop searching; can't name a verification → ask one pointed question). The full method is skills/fable-method/SKILL.md, ~110 lines, every sentence load-bearing.

The whole thing, as one flowchart