Skip to content
Solid State
Solid State Original · Write-mode audit

Your setup,
ready for Fable 5.

Fable 5 changed the rules underneath Opus-era setups. Instructions that helped now hurt. Prompts that worked now end in refusals that look like success. fable-ready scans your repo, patches what you approve, and stamps a dated verdict.

> fable-ready --target . --apply ask

F1  skills/research/SKILL.md:12   "show your full reasoning"   blocker
    PATCH → "state conclusions, list the evidence"             [apply? y]
F5  agents/researcher.json:4      "budget_tokens": 8000        blocker
    PATCH → remove; Fable rejects manual thinking budgets      [apply? y]

Re-scan: 0 blockers.
VERDICT: FABLE-READY (2026-06-12) → FABLE-READY.md
once · rules updates through v1.x · zip → unzip → verdict
01 — What breaks

Five families. Sourced, not vibed.

Every rule carries an evidence grade and a source — an Anthropic doc, or a flagged community report. The worst failures return HTTP 200 and look like success. That's why this is a scan, not a checklist.

F1 · Reasoning-display triggers“Show your reasoning” prompts hit a refusal category that returns HTTP 200 — your pipeline reads it as success.blocker
F2 · Prescriptive step-listsOpus-era leashes degrade Fable output. Anthropic's own guidance: rewrite to intent + boundaries.blocker
F3 · Missing output disciplineFable narrates at high effort. No brevity rule means paying output prices for prose you delete.warn
F4 · Missing boundariesFable acts on initiative — drafts the email, cuts the branch. No do/don't block means surprises.warn
F5 · Mechanics driftManual thinking budgets now 400-error. Stale model strings run the old model silently. Timeouts tuned to Opus turns break mid-run.blocker

Also in the zip: effort-map.md (task type → effort level, with the cost math), long-run.md, memory-bootstrap.md.

02 — How it runs

Scan. Patch. Verdict.

scan.py walks CLAUDE.md, skills, commands, agent configs — deterministic, stdlib, no network
the model drafts the exact rewrite per finding, from the playbooks
you approve each patch — meaning-changing patches are never auto-applied
re-scan → FABLE-READY.md: dated verdict, rules version printed
.solidstate/ present? The verdict logs itself. Ship Kit interop, not dependency.

The scanner catches what regexes can catch — reproducibly. The model judges what regexes can't. A flagged line is a question, not a conviction.

03 — Not for everyone

Skip this if:

You don't have a CLAUDE.md or custom skills yet. There's nothing to audit — start with the free skills.

You only use chat. Write-mode patches need a file-capable tool: Claude Code, Cowork, Cursor, OpenClaw, Hermes. Paste-only runs report-only.

You want a spend meter. This audits files, not bills — the cost arithmetic ships as a reference, the dashboard doesn't.

04 — Provenance

Fable 5 patched its own migration.

Written and audited by Claude (Fable 5) on the production line behind solidstate.cc — validated, positioned, audited, and launched with Ship Kit, on camera. The scanner ran on its own fixtures before it shipped; the example report in the zip is real output. The full run is published as a case study, receipts included.

Dated honesty: the rules are graded against the Fable 5 docs of 2026-06. Rules move; v1.x updates are included, and every report prints its rules version.

$49
once · rules updates through v1.x · no subscription
Stripe checkout → instant zip download. One purchase, one operator — same honest terms as the Ship Kit.

Shipping a product, not just migrating one? The Ship Kit is the system this skill was built with.