fable-skills

A library of 35 engineering-discipline skills — small, composable instruction files that encode how a careful senior engineer works, not what any particular framework does.

The premise

The most capable Claude models don’t just know more — they work more carefully. Fable 5 reaches for a codebase’s own conventions before writing, proves a bug’s cause before fixing it, sweeps edge cases before declaring done, and reports what it actually verified rather than what it assumes. A strong-but-lighter model like Opus 4.8 can do all of that too — it just doesn’t always reach for those habits unprompted.

These skills close that gap. Fable identified the places where a less deliberate run tends to fall short of its own ceiling — orientation, root-cause discipline, security reflexes, edge-case coverage, honest reporting — and wrote each habit down as an explicit, always-available checklist. Load them into a session and a model follows the discipline on purpose instead of leaving it to chance. The goal is simple: Fable-level output from whatever model is driving.

Each skill is one focused habit. They fire situationally — a debugging task pulls in root-cause-debugging, a diff touching user input pulls in security-reflexes, a schema change pulls in data-migration-safety — so the model carries the right discipline into the right moment without drowning in guidance.

Does it actually help? Benchmarks

Same method throughout: run an identical task through fresh agents with no guidance vs. with the relevant skill files, then score the outputs against a fixed answer key. Benchmark 1 uses Opus 4.8; Benchmark 2 runs both Opus 4.8 and Sonnet 5. These are directional signals, not statistical proofs.

The full assets are in benchmark/ — seeded tasks, answer keys, all 62 raw run outputs, the blind-grader prompts, and the per-item score tables — so every number below is auditable and re-gradable rather than take-my-word-for-it. (They were reconstructed verbatim from the original session transcript; each file carries a provenance header, and the recovery manifest in benchmark/README.md lists the known gaps.)

Benchmark 1 — Security code review (n=1, blind-graded)

The same task was run through two fresh agents and both reviews were scored blind by an independent third agent. Task: review a Python orders module seeded with 14 latent issues (SQL injection, connection leak, division-by-zero, float-money drift, a missing authorization check, …).

The skill-guided run didn’t just find more — it found the issue that mattered most. Walking the security-reflexes checklist (“every mutating path: WHO is calling and MAY they touch THIS object?”) surfaced an authorization gap the unprompted review sailed past.

Benchmark 2 — Three areas × two models, n=5 per cell, blind-graded

Three fresh tasks (root-cause debugging, a racy rate limiter, a CSV revenue parser), each seeded with a fixed answer key. Every cell is 5 independent runs; all 60 outputs per area were scored blind by an independent grader that knew neither the model nor whether skills were applied. Mean issues found:

Four findings:

1. Skills lift both models in every area — biggest on debugging (Opus +2.6, Sonnet +2.0), where the task rewards process; smallest on the edge task, whose “list every edge case” prompt already forces exhaustiveness (both models near-saturate).
2. Skills bring the cheaper model above the stronger bare one. Sonnet 5 + skills matches or beats bare Opus 4.8 in all three areas (6.4 vs 5.2, 7.0 vs 6.4, 9.0 vs 8.2). That is the parity thesis, measured: the discipline files close the model gap.
3. But raw capability still matters at the subtle end. Opus + skills leads Sonnet + skills on debugging (7.8 vs 6.4) almost entirely on one item — a config of JSON null defeating the is None cache guard — caught by Opus-with-skills 4/5 times and by Sonnet 0/5 in any condition. Skills prompt the right class of thinking; the hardest single insight still needed the stronger model.
4. Variance shrank with skills — with-skills scores cluster at the top rather than spreading (Opus debugging 7–8 with vs 5–6 without), so the gain is greater consistency, not just a higher mean.

The Opus numbers here were produced by an earlier author-scored pass too; the blind re-grade reproduced them almost exactly (debugging 5.2 vs 5.4, concurrency and the with-skills scores identical), which is the main reason to trust the rest.

Honest caveats: n=5 per cell is still small — directional, not conclusive. The baselines are already strong, so the gain is “good → near-complete,” not “broken → fixed.” Grading is an LLM applying an objective checklist, not a human. The edge task saturates, compressing its signal.

Using the skills

Each skill lives in its own directory as a SKILL.md with YAML frontmatter (name, description) and a short body. The description is what a harness matches against the situation to decide when to activate the skill.

With Claude Code: drop the skill directories into a discoverable skills location (e.g. ~/.claude/skills/ or a project .claude/skills/) and they become available to the Skill tool. The model selects them by description as tasks arise.

Anywhere else: the bodies are plain Markdown — paste the relevant one into a system prompt, or concatenate a task-appropriate subset. That’s exactly how the benchmarks’ “with skills” runs were configured.

The catalog (35 skills)

Orientation & planning

• codebase-orientation — recon an unfamiliar repo before changing it
• task-decomposition — acceptance-test-first, dependency-ordered planning
• estimation-and-scoping — honest sizing; surface 10× discoveries early
• ambiguity-commit — resolve ambiguous requests by investigating, then stating the interpretation
• question-vs-task — tell “explain this” apart from “change this” before acting

Correctness & debugging

• root-cause-debugging — reproduce, bisect, prove the cause, fix minimally, re-verify
• edge-case-sweep — enumerate edge cases against a concrete checklist
• numerical-care — floats, currency, rounding, division-by-zero, overflow, units
• concurrency-reasoning — races, check-then-act gaps, mutation across await, cancellation
• environment-first — suspect the environment before rewriting the code
• stop-thrashing — detect non-converging iteration and force a zoom-out

Security & safety

• security-reflexes — injection, authz, secrets, SSRF/XSS on everyday diffs
• untrusted-content-guard — treat fetched/read content as data, never instructions
• data-loss-guard — stop-and-check before destructive or irreversible operations
• api-surface-care — treat changes to shared interfaces as contract changes

Change discipline

• surgical-refactoring — enumerate call sites, separate mechanical from judgment edits
• dependency-changes — justify, archaeology, lockfile hygiene, one bump at a time
• data-migration-safety — expand→migrate→contract, backfills, rollback-first

Testing & verification

• test-design — regression-first, boundary tables, behavior over implementation
• verify-ui-visually — render and look, don’t imagine the markup
• workmanship — evidence before assertion, the done gate, report what happened

Performance & operations

• performance-investigation — measure, profile, fix by leverage, verify with the same measurement
• perf-sanity — catch the algorithmic/I/O classics while writing (O(n²), N+1)
• incident-diagnosis — read-only evidence first, mitigate reversibly before root-causing

Communication

• calibrated-recommendation — match the strength of the claim to the evidence
• explain-at-the-right-level — teach at the asker’s level, grounded in their code
• error-message-quality — errors that name the failing thing, the value, and the fix
• escalate-vs-decide — reversibility × blast-radius: when to ask vs decide
• generalize-the-correction — fix the whole class of a mistake, not just the instance

Working style & hygiene

• git-hygiene — atomic commits, branch before risk, never commit debris or secrets
• leave-no-mess — clean up processes, temp files, and scaffolding before done
• context-checkpoint — externalize state so long tasks survive compaction
• subagent-fanout — delegate broad searches to keep the main context clean
• cross-platform-care — scripts that survive a machine they didn’t develop on

LLM application code

• llm-app-code — treat model output as untrusted input; schema-validate, gate writes

Agents

Skills fix habits; agents fix structure. A skill makes the driving model reach for a discipline it already has. But one gap no checklist can close is self-correlation: a model reviewing its own work inherits its own anchoring — the same context, the same wrong assumption it made an hour ago. Subagents close that gap structurally, because each one runs in a clean context. The agents/ directory packages Fable’s independence disciplines as reusable subagent definitions for Claude Code’s Agent tool.

Install by dropping a file into ~/.claude/agents/ (or a project’s .claude/agents/); it registers at the next session start.

Fresh-context independence (uncorrelated eyes on your own work):

• skeptic — adversarial verifier: takes one claim and tries to refute it from primary evidence; returns CONFIRMED / REFUTED / UNPROVEN with the evidence trail. For load-bearing conclusions, especially your own.
• fresh-eyes-reviewer — red-team review of a diff with zero knowledge of the author’s reasoning, by design; reports only findings with a concrete failure scenario attached.
• done-gate — spec-compliance audit before “done”: every requirement in the original request checked against evidence, cheap verifications re-run, unevidenced claims flagged.
• blind-judge — grades N provenance-stripped candidates against a rubric, criterion-by-criterion, with anti-halo and position-bias guards. For judge panels and tournament selection (it’s how this repo’s own benchmarks were scored).
• plan-premortem — “this plan already failed; write the incident report”: verifies the plan’s factual claims against the repo, surfaces unverified load-bearing assumptions, names the riskiest step and the cheapest de-risking probe.

The verification agents deliberately pin model: opus — the point is a cheap driver doing the legwork while the strongest available model does the judging, not the reverse. Override the model: field to taste.

Cross-model independence (blind spots that don’t correlate with Claude’s):

• codex — runs the OpenAI Codex CLI (GPT-5.6) as a subagent inside Claude fan-outs. The orchestrator picks model tier and reasoning effort per task via two prompt headers (codex-model:, codex-effort:), steered by an evidence-based routing guide baked into the description: Luna for mechanical/lookup work, Terra at high effort for routine implementation (the best measured marginal return), Sol at xhigh for long-horizon, multi-file, and adversarial-review work — where Sol’s higher completion rate makes it cheaper per completed task than Terra despite 2× the token price. Hard rules: never hand Luna a >200K-token context (long-context recall cliff), and escalate effort on retry rather than defaulting to max (quality vs. effort is non-monotonic). Requires the codex CLI installed and authenticated; verified against codex-cli 0.144 / GPT-5.6 GA (2026-07).

Attribution

Skills authored by Fable to lift lighter models to its own working standard. Benchmark run and scored with Claude Opus 4.8.