Skip to content

LewenW/make-opus-fable

Repository files navigation

🜂 make-opus-fable

Translate Fable 5's trained-in work habits into explicit process + structure Opus 4.8 can execute. Turn Claude Opus 4.8 into a more Fable-like operator — a comprehensive discipline layer across the whole working loop: verify, audit, debug, write native, plan, ground, judge.

License: MIT For Claude Code Model: Opus 4.8 Skills: 9 Evidence: 13 batches

An evidence-backed suite of behavior / orchestration skills, built on the most exhaustive public Opus↔Fable eval sweep we're aware of — 13 batches of objective three-arm testing (bare Opus / Opus+suite / Fable) across nearly every axis that could conceivably separate them, including conclusions our own held-out testing later overturned. The result: near-total parity everywhere, and this suite closes most of what's left.


⚡ One-command install

git clone https://github.com/LewenW/make-opus-fable.git
cd make-opus-fable && bash install.sh

Start a new Claude Code session for it to take effect. Try it right away:

/verify-before-done      pre-flight check before declaring anything done — catches "looks right but isn't"
/deep-audit              audit a whole repo for bugs (per-file fan-out reviewers + xhigh, buys recall)
/quant-thesis            forecast a downstream number from messy upstream signals, with shown arithmetic and two-tier conviction
Command What it does
bash install.sh Installs 9 skills + the verifier subagent + the behavior discipline block
bash install.sh --with-hooks Also installs two deterministic hooks: verify-after-edit (runs your project's tests after every code edit) and deep-audit-trigger (makes the audit fan-out fire reliably)
bash install.sh --uninstall One-command removal (manages CLAUDE.md via a marked block, never touches your existing content)

Idempotent, safe to re-run. Installs to the user-level ~/.claude/; the behavior core is appended to ~/.claude/CLAUDE.md inside a block marked make-opus-fable, and precisely removed on uninstall. Tested end to end: install → reinstall → uninstall leaves your original content untouched.

Or install as a Claude Code plugin

/plugin marketplace add LewenW/make-opus-fable
/plugin install make-opus-fable@make-opus-fable

The plugin path delivers the 9 skills + the verifier agent (namespaced, e.g. /make-opus-fable:deep-audit) with zero cloning. Because plugins can't modify your CLAUDE.md or auto-load these hooks, the always-on behavior core and the two deterministic hooks still come from install.sh — run it if you want the full suite. Use whichever entry point fits; they don't conflict. (Marketplace manifest passes claude plugin validate.)


🎯 What it covers, and how we know it works

This is a comprehensive engineering-discipline suite, not a niche add-on. Nine skills + two hooks cover the whole working loop — verify before done, audit for defects, debug to root cause, write code native to the file, keep scope tight, plan long work, ground visual output, form quantitative theses, make design calls, and carry memory across sessions. The goal is a general operator that behaves like a careful senior engineer across everything, not a specialist in one domain.

The quant/finance angle is not the product's identity — it was one of the probes we used to measure where Opus actually lags Fable. Which brings us to the evidence:

We didn't guess where the gaps were — we swept for them. 13 batches of three-arm evals (bare Opus / Opus+suite / Fable) covered nearly every axis that could separate the two: self-checkable coding, terminal-agentic repair, instruction-following (up to 85 concurrent constraints), long-context multi-hop retrieval, knowledge density, honesty under missing context, financial calculation, quant forecasting, judgment calls, and exhaustive code audits.

The result across that whole sweep: near-total parity. On coding, terminal repair, instruction-following, retrieval, knowledge, and financial calculation, the two models are indistinguishable — both at ceiling — and the suite's discipline layer holds Opus there with no drag. Real, measured gaps showed up on three axes, and the suite closes most of what's closable:

Axis Gap before the suite With the suite
Quant thesis / forecasting (forming a call from messy data) Opus lost the vote 6:27 15:18 — 86% of the gap closed (quant-thesis)
Exhaustive defect recall (finding every bug in an audit) Opus 5.5/10 ~43% closed, near Fable on real production code (deep-audit)
Behavioral discipline (honest reporting / no fabrication / scope control) Fable more consistent Effectively closed (blind eval: 15 wins, 2 losses)
Visual perception / raw knowledge density Fable stronger ❌ Not closeable by any suite — baked into model weights

That's the whole picture: an exhaustive, evidence-first sweep, near-total parity everywhere a clean task lets the two models compete fairly, and full transparency about the one place (perception) that no amount of structure or prompting can fix. We looked — there's no comparably broad, held-out-validated Opus↔Fable eval sweep we're aware of in any public skill suite. What we won't do is claim to be "the closest to Fable on the market": we haven't run a head-to-head against every other project out there, and we'd rather under-claim than repeat the mistake this suite's own eval history caught and corrected (see evals/HARDBENCH.md — an earlier "beats Fable" claim from batch 4 was overturned by held-out testing in batch 6).


🧩 Nine skills

Once installed, each can be triggered manually with /<name>, or automatically based on its description (skills are known to under-trigger — manual invocation is recommended for anything that matters).

Skill When to use it What it does
🔍 verify-before-done Before declaring any substantive work "done / fixed / passing" Evidence audit → adversarial five-point pass → only findings go into the deliverable, never the process narration. Kills "looks right but isn't"
🗂 deep-audit The goal is to find every defect: pre-merge multi-file audits, "review this module for bugs", regression / security passes Enumerate files → one fresh xhigh reviewer per file, in parallel → union, then de-dupe and re-verify. Trades tokens and wall-clock for coverage
📈 quant-thesis Forecasting a downstream number from upstream signals: "what does X imply for Y", "will revenue accelerate", "read-through" Shows the decomposition arithmetic explicitly, sizes pass-through with coefficients, gives a numbered band, splits conviction into direction vs. magnitude, checks base effects / stock-vs-flow
⚖️ judgment The deliverable is a decision / design / assessment, not an edit: "should we do X or Y", "is this design sound" Assess before editing, lead with the call plus its one real tradeoff, run a blindspot pass, don't implement until agreed
🧭 long-horizon-protocol Spans multiple files / steps / sessions: refactors, migrations, whole features, cross-module debugging, long research Consolidate requirements → plan-gate → slice into ≤1h units → evidence ledger + completion gate → checkpoint state
🧠 memory-discipline Reading or writing cross-session memory: CLAUDE.md, progress notes, lesson ledgers What to write / how to write it / verify before recalling — never carry an assumption into a future session as if it were fact
🖼 visual-grounding The deliverable is visual: rendered HTML/CSS, SVG, charts, UI, animations, any layout/styling change Render it and look at the actual output (screenshot / DOM / computed styles) before claiming it works — a passing build verifies code, not pixels
🐛 debugging Locating the cause of one specific failure: "why is X failing", a failing test, a crash, a wrong output Reproduce deterministically → hypothesis with a falsifier → binary-search / git bisect to isolate → confirm the causal chain → minimal fix → re-run the repro. Never guess-and-retry
🧵 native-code Writing or editing code inside an existing file Match the file's idiom; no defensive bloat, no narrating comments, no speculative generality — the change should read as if the original author wrote it

Plus agents/verifier.md — a fresh-context adversarial verification subagent. Invoke @verifier by name on high-stakes changes to guarantee the check actually runs (stronger than self-review). And two deterministic hooks (--with-hooks): verify-after-edit runs your project's tests after every code edit (debounced, with a timeout and env-var kill switch), deep-audit-trigger guarantees the audit fan-out fires — a skill is a request, a hook is a guarantee.


🔬 Why it works (the mechanism, in one line)

The gap isn't that Opus "can't" — it's that it hasn't built the reflex.

  • Auditing: Opus can find bugs, but doesn't reach for exhaustively enumerating every file on its own. → deep-audit uses fan-out orchestration to turn "breadth" into N in-window units — recall climbs from 5.5/10 on bare Opus to near Fable.
  • Quant: Opus can compute 1.14 / 1.08, but doesn't reach for showing the decomposition on its own. → quant-thesis codifies Fable's technique reflexes into a protocol, closing 86% of the vote-level gap.
  • Capability axes (vision / knowledge): these are baked into the weights, and no prompt can add them — the suite is honest that it can't help here.

So the levers that actually work are configuration (effort) + structure (fan-out orchestration) + targeted reflex protocols — not "writing a smarter prompt." Plain-text skills, tested against held-out cases, repeatedly close 0% of the capability-axis gap.


📊 Evidence

Every conclusion (including the ones held-out testing overturned) lives in evals/HARDBENCH.md — 13 batches of objective three-arm evals, graded by hidden tests, on-disk state, Python, or blind pairwise panels. Highlights:

  • 6 self-checkable coding tasks → all three arms hit 100%, zero gap (clean single-turn tasks never separate the two models).
  • Defect recall → bare Opus 5.5/10; xhigh + fan-out closes most of the gap; on real production code under the same protocol, head-to-head: Opus 7/9 vs. Fable 9/9 (closer, not closed).
  • Quant thesis (n=11, blind pairwise panel) → bare Opus loses to Fable 6:27 on votes → quant-thesis closes it to 15:18, nearly even (86% closed).
  • Four capability-axis follow-ups (terminal agentic / instruction-following at 85 concurrent constraints / long-context multi-hop / knowledge density) → all four axes show zero separation across all three arms — every one at ceiling.
  • Real-world validation → a full fan-out audit of a production repo surfaced 1 critical + 8 high-severity bugs, all confirmed real (evals/AUDIT-PARALLAX.md).

New here? evals/SHOWCASE.md is the 2-minute before/after version — three concrete flips (quant 6:27 → 15:18, defect recall 5.5 → near-Fable, behavior 15-2) without the full eval density.


🛠 Usage notes

  • Auto-triggering isn't fully reliable (a known Anthropic limitation). For anything that matters, invoke manually: /long-horizon-protocol, /verify-before-done before declaring done, @verifier for review.
  • Don't apply process to small tasks — every skill has a trivial-task escape hatch; if the overhead feels noticeable, that's a sign the trigger scope should be narrowed.
  • Token cost tiers (measured): generation / judgment / everyday work ≈ ; long-horizon ≈ 1.2–1.5×; deep-audit fan-out ≈ 4–5× (the one real cost center — only pay it when you actually want audit-grade recall).
  • For quant / audit work, pair with effort=xhigh (already built into deep-audit's frontmatter).

📁 Layout

skills/
  verify-before-done/     pre-done check: evidence audit → adversarial pass → findings in, process out
  deep-audit/             fan-out audit: effort xhigh (frontmatter) + per-file reviewer + de-dupe re-verify
  quant-thesis/           quant reflexes: show decomposition arithmetic / sized pass-through / numbered bands / two-tier conviction
  judgment/               decisions/design: assess before editing, lead with the call + one tradeoff, blindspot pass
  long-horizon-protocol/  long tasks: consolidate → plan-gate → slice → evidence ledger + completion gate → checkpoint
  memory-discipline/      memory hygiene: what/how to write, verify before recalling
  visual-grounding/       visual work: render and look at the actual output before claiming it works
  debugging/              root-cause: reproduce → isolate → causal chain → minimal fix → re-run repro
  native-code/            match the file's idiom; no defensive bloat, narration, or speculative generality
agents/verifier.md        fresh-context adversarial verification subagent (@verifier = guaranteed execution)
config/
  CLAUDE-core.md          persistent behavior core + task router (install into ~/.claude/CLAUDE.md; 5 modes)
  settings-snippets.md    config levers: thinking / effort / hooks / API parameters
hooks/                    two deterministic backstops: verify-after-edit (PostToolUse) + deep-audit-trigger (UserPromptSubmit)
.claude-plugin/           plugin + marketplace manifests (install via /plugin marketplace add)
evals/
  SHOWCASE.md             2-minute before/after highlights (start here)
  HARDBENCH.md            13 batches of objective three-arm evals (core evidence; includes overturned conclusions)
  RESULTS.md              full blind-eval track record and iteration history (15-2-0; includes every loss, diagnosis, lesson)
  AUDIT-PARALLAX.md        full fan-out audit report on a production repo (1 critical + 8 high, all confirmed real)
install.sh                one-command install / uninstall (idempotent, reversible)

❓ FAQ

Will it slow down simple tasks?

No. Every skill has a trivial-task escape hatch; generation / judgment / everyday overhead ≈ 1×. Measured: on minute-scale, self-checkable tasks, the suite adds zero accuracy gain and zero drag.

Will it touch my existing CLAUDE.md?

No. The behavior core is appended inside a block marked make-opus-fable; --uninstall removes exactly that block, leaving your original content untouched. Verified by testing.

Where does it install? Can it be project-scoped?

By default it installs to the user-level ~/.claude/ (global). To scope it to one project, copy skills/* into that project's .claude/skills/ instead.

Does this turn Opus into Fable?

No — honestly. It gets Opus "close enough" to Fable on the three axes where a real gap exists (quant 86% / recall 43% / behavioral discipline effective), but it doesn't fully close the gap — perception and raw knowledge are baked into training, and no prompt fixes that. Arguably the most valuable part of this project isn't the skills themselves, but the rigorous, self-skeptical eval discipline behind them (held-out validation, isolated variables, objective oracles, spot-checking).


🙏 Prior art & acknowledgments

This suite aims to be comprehensive, and part of getting there was studying the other open Fable-behavior projects and folding in the general engineering disciplines they got right (ideas, not code — each is independently authored here). Credit where due, all MIT-licensed:

  • fivetaku/fablize — the most rigorous of the bunch; its verification-grounding and root-cause investigation framing informed debugging and visual-grounding.
  • DizzyMii/fable-skills — the "write native to the file" discipline behind native-code, and the rationalization-and-rebuttal format.
  • benjaminard/fable-skills — root-cause-first and minimal-diff, which reinforced debugging and the scope discipline in the core.
  • alicicek/tale-mode — deterministic gate hooks; a reminder that a hook is a guarantee where a skill is only a request.

Where these projects converge — verify before done, fan-out review, finish the turn, outcome-first — we converge too; the shared conclusion is the point. What's distinct here is the evidence chain: 13 batches of objective three-arm evals (including conclusions our own held-out testing overturned), rather than assertion. We deliberately do not ship the more intrusive hooks some of these use (turn-end blocking, test-vs-live-data guards) — that's a design choice toward advisory-by-default, not an oversight.


Uninstall · bash install.sh --uninstall

Built with honest evals. MIT licensed.

About

Make Opus 4.8 more Fable-like: discipline, audit breadth, and quant reflexes — 6 skills backed by 13 batches of honest three-arm evals.

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors