anthropic-grade-optimizer

Cite or stay silent. Artifact type comes first. Voice drift trumps score. An auditor for Claude-directing artifacts against 189 cited Anthropic rules across 11 dimensions, calibrated per model. Every finding ships with verbatim quote and source URL. Authorial voice is preserved by contract — not by accident.

Why this exists

Most prompt linters score what's easy to count and call it a day. This one starts from a different premise: a Claude-directing artifact is a cognitive contract, not a configuration file. Audit it like a contract — cite the clause, name the violation, propose the patch, and never rewrite the author into someone else. If the auditor cannot point to the doctrine, the auditor stays silent.

That is the line this skill refuses to cross, and the reason it audits itself on every release.

The Three Laws

Every finding, every diff, every score lives under these three rules. Violating any one collapses the others.

1. Cite or stay silent. Every 🔴 / 🟡 finding carries a source_url and a verbatim quote. No source → no finding.
2. Artifact type comes first. A SKILL.md rule fired against a CLAUDE.md is a false positive, not a finding. The rubric is type-aware before it is severity-aware.
3. Voice drift trumps score. An optimization that lifts the score by diluting the operator's voice is a regression in disguise. Drift over 10% aborts; with --push-ceiling the gate tightens to 5%.

What it does

Reads any Claude-directing artifact — CLAUDE.md, SKILL.md, subagent, hook, MCP config, system prompt, user prompt, workflow, or api_config — and emits:

• a 0–100 score with letter grade
• per-dimension breakdown across the 11 dimensions (D-CLAR, D-STRUCT, D-EXAMPLE, D-REASON, D-CONTEXT, D-MODEL, D-AGENT, D-EVAL, D-TOOL, D-VISION, D-CC for the Claude Code ecosystem)
• triaged findings (🔴 must-fix, 🟡 should-fix, 🟢 may-fix, ❓ open question, ⚪ preserve)
• surgical diffs (mode ≠ audit), each citing the rule that triggered them
• voice-fidelity gate (drift > 10% aborts optimization)
• declared-exception ledger with side-by-side raw/active scores

Install

As a Claude Code skill (recommended):

git clone https://github.com/l0z4n0-a1/skill-anthropic-grade-optimizer.git \
  ~/.claude/skills/anthropic-grade-optimizer
cd ~/.claude/skills/anthropic-grade-optimizer
pip install pyyaml                 # required
pip install anthropic              # optional, only for --pass2

The skill auto-loads in any Claude Code session. Trigger it by asking Claude to "audit this CLAUDE.md", "make this skill Anthropic-grade", "is this prompt good for Opus 4.7?", etc.

As a standalone CLI (no Claude Code required):

git clone https://github.com/l0z4n0-a1/skill-anthropic-grade-optimizer.git
cd anthropic-grade-optimizer
pip install pyyaml
python scripts/run.py path/to/SKILL.md --target opus-4-7 --mode audit

Requirements: Python 3.10+. pyyaml is the only mandatory dependency. The anthropic SDK is only needed for the --pass2 qualitative-judge mode and degrades gracefully when absent.

Quick start

python scripts/run.py <artifact_path> --target opus-4-7 --mode audit

With Pass-2 qualitative audit (requires ANTHROPIC_API_KEY):

python scripts/run.py <artifact_path> --target opus-4-7 --pass2

Strict mode (no operator-declared exceptions):

python scripts/run.py <artifact_path> --target opus-4-7 --no-exceptions

Generate diffs and write to a sibling <artifact>.optimized.<ext>:

python scripts/run.py <artifact_path> --target opus-4-7 --mode optimize --apply

How it's verified

Layer	What it covers	Run
Strange-loop self-audit	Skill audits its own SKILL.md every release	python scripts/run.py SKILL.md --target opus-4-7
Eval suite	15 fixtures across good / bad / adversarial / cross-bad	python scripts/run_eval_suite.py
Unit tests	Rule-schema integrity (one test per rule_id), detector output shape, orchestrator contract	python -m pytest tests/ -p no:langsmith_plugin
Calibration	Precision / recall / F1 vs ground truth	python tests/calibration.py
Model currency	CI fails when modulation-matrix drifts from known-current models	python scripts/check_model_currency.py
CI	All of the above on every PR	.github/workflows/ci.yml

Honest scope

• What it does: audits prompts and Claude-directing artifacts against published Anthropic doctrine.
• What it does not: evaluate Claude's responses (use eval pipelines), or audit application code that consumes the API beyond the prompts themselves.
• Doctrine limits: see references/gaps.md for declared coverage gaps. Anthropic doctrine is prescriptive, not diagnostic; the 11-dimension rubric is auditor-canonical, not Anthropic-canonical.

Coverage that cannot be cited is held back, not invented. Silent gaps are worse than declared ones — so we declare them.

A note on the skill's own name

The skill is named anthropic-grade-optimizer even though Anthropic's own SKILL guidance reserves the word anthropic in skill names. This is a declared exception registered in scripts/run.py § DECLARED_EXCEPTIONS: the name is the most precise descriptor of the skill's purpose, the rule continues to fire on every self-audit, and the orchestrator always shows both active (with the exception) and raw (--no-exceptions) scores side by side. Voice fidelity over hidden adjustment — see evals/SELF-AUDIT.md.

Architecture

SKILL.md                     ← canonical entry point
scripts/
  run.py                     ← orchestrator: classify → pass1 → [pass2] → score → emit
  classify_artifact.py       ← detect type + signals
  pass1_mechanical.py        ← regex / code-check / heuristic detectors
  pass2_executor.py          ← Claude API qualitative judge with prompt caching
  score_calculator.py        ← per-dimension aggregation + voice-drift estimate
  optimize.py                ← surgical diffs + snippet patches
  check_model_currency.py    ← CI guard against modulation-matrix drift
  run_eval_suite.py          ← replay 15 ground-truth fixtures
references/
  rules-anthropic.yaml       ← 189 rules with verbatim_quote + source_url (SSOT)
  rules-master.yaml          ← dimension index
  modulation-matrix.yaml     ← 4 model profiles
  rubric-by-type.yaml        ← rule subsets per artifact type
  pass2-protocol.md          ← qualitative audit protocol
  gaps.md                    ← declared doctrinal limits + 5 resolved conflicts
  anti-patterns.md           ← AP-1..15
  heuristics.md              ← H1..12
  quotes-canonical.md        ← citation library
assets/snippets/             ← 18 verbatim canonical insertions for high-leverage rules
tests/
  test_rule_schema.py        ← one test per rule_id
  test_pass1_detectors.py    ← detector output well-formedness per fixture
  test_orchestrator.py       ← bundle contract, alias compat, exception toggle
  calibration.py             ← precision / recall / F1 vs ground truth
evals/
  fixtures/                  ← 15 ground-truth inputs
  ground-truth.yaml          ← expected scores + grades
  SELF-AUDIT.md              ← committed strange-loop result
  results/                   ← auto-generated EVAL_MATRIX.md, CALIBRATION_REPORT.md

Contributing

See CONTRIBUTING.md. The short version: every new finding needs a verbatim Anthropic quote with source URL, and every new rule must include a fixture and a unit test.

Security

See SECURITY.md for vulnerability reporting.

License

MIT — see LICENSE.

---

Shipped under the discipline it audits. The strange-loop closes cleanly: this skill audits itself with the same pipeline it audits everything else, and the result is committed at evals/SELF-AUDIT.md on every release.

Identity before architecture. Coherence before completeness. Form demonstrates content.