Codebase Analyzer

Codebase Analyzer Skill

Statistical rule discovery through measurement of Go codebases. Python scripts count patterns to avoid LLM training bias, then statistics are interpreted to derive confidence-scored rules. The core principle is Measure First, Interpret Second -- what IS in the code is the local standard, not what an LLM thinks "should be" there.

Deep References

Load on demand when the corresponding signal appears.

Instructions

Phase 1: CONFIGURE

Goal: Validate target and select analyzer variant.

Read and follow the repository's CLAUDE.md before doing anything else -- project instructions override default behaviors.

Step 1: Validate the target

• Confirm path points to a Go repository root with .go files
• Check for standard structure (cmd/, internal/, pkg/)
• Verify sufficient file count: 50+ files for meaningful rules, 100+ ideal. Below 50 files, statistics produce high variance -- patterns that look consistent may be coincidence. For small repos, combine analysis across multiple team repos rather than treating thin data as definitive.

Step 2: Select cartographer variant

Step 3: Verify environment

• Python 3.7+ available
• No external dependencies needed (uses only Python standard library)
• Output directories exist or can be created

See references/phase-details.md for the CONFIGURE banner template.

Gate: Target directory exists, contains 50+ Go files, variant selected. Proceed only when gate passes.

Phase 2: MEASURE

Goal: Run statistical analysis scripts. Pure measurement -- no interpretation yet.

This phase is strictly mechanical. Scripts count and measure; keep interpretation separate from data collection. Combining measurement with interpretation introduces LLM training bias -- the model reports what "should be" instead of what IS. Run scripts first, interpret the numbers second, always as separate steps.

Automatically filter vendor/, testdata/, and generated code (files with "Code generated by..." markers) to avoid polluting statistics with external patterns.

Step 1: Execute the cartographer

bashCopy code
python3 ${CLAUDE_SKILL_DIR}/scripts/cartographer_omni.py /path/to/go/repo
# Or for quick overview: python3 ${CLAUDE_SKILL_DIR}/scripts/cartographer.py /path/to/go/repo

Always run the cartographer scripts for measurement; reserve LLM interpretation for Phase 3. When an LLM sees return err it may report "not wrapping errors properly" even if that IS the local standard. The scripts produce deterministic, reproducible counts; the LLM's role begins at interpretation in Phase 3.

Step 2: Verify output integrity

• Confirm JSON output is valid and complete
• Check file count matches expectations (no vendor pollution)
• Verify all three lenses produced data
• Confirm derived_rules section exists in output

Step 3: Check for data quality issues

• File count suspiciously high? Vendor code may be included
• File count suspiciously low? Subdirectories may be missed
• All percentages near 50%? May indicate mixed codebase or insufficient data

See references/phase-details.md for the MEASURE banner template.

Gate: Script completed without errors, JSON output is valid, file count is reasonable. Proceed only when gate passes.

Phase 3: INTERPRET

Goal: Derive rules from statistics. This is where LLM interpretation happens -- AFTER measurement is complete.

Report facts and show complete statistics rather than describing them. Report facts without editorializing about code quality -- the numbers speak for themselves.

Step 1: Review the three lenses

For detailed lens explanations, see references/three-lenses.md.

Step 2: Extract rules by confidence

Only derive rules from patterns with sufficient consistency. Forcing rules from weak patterns causes false positives in reviews and may impose standards the team has not organically adopted.

Step 3: Review Style Vector (Omni only)

• 10 composite scores (0-100): Consistency, Modernization, Safety, Idiomaticity, Documentation, Testing Maturity, Architecture, Performance, Observability, Production Readiness
• Identify strengths (scores >75) and gaps (scores <50)
• Note shadow constitution entries (accepted linter suppressions)

Step 4: Cross-reference lenses

• Pattern confirmed across multiple lenses = higher confidence
• Pattern in one lens only = standard confidence
• Contradictions between lenses = investigate further

Gate: Rules extracted with evidence and confidence levels. Style Vector reviewed. Proceed only when gate passes.

Phase 4: DELIVER

Goal: Produce actionable output artifacts.

Step 1: Save statistical report

Copy code
cartography_data/{repo_name}_cartography.json

Step 2: Generate derived rules document

Copy code
derived_rules/{repo_name}_rules.md

Rule and Style Vector formats, plus the DELIVER banner template, live in references/phase-details.md.

Step 3: Summarize Style Vector (Omni only) — see phase-details.md

Step 4: Recommend next steps

• Compare with pr-workflow (miner) data if available (explicit vs implicit rules)
• Suggest CLAUDE.md updates for high-confidence rules
• Identify golangci-lint rules that could enforce discovered patterns
• Suggest quarterly re-analysis schedule -- coding patterns evolve with team growth and new Go versions, so a one-time snapshot becomes stale within months

Gate: JSON report saved, rules document generated, next steps documented. Analysis complete.

Complementary Skills, Examples, Error Handling

Load references/phase-details.md for:

• Complementary skills (pr-workflow miner) and reconciliation matrix
• Worked examples: single repo, team-wide discovery, onboarding
• Error catalog: no Go files found, no rules derived, vendor/generated pollution

Prerequisites

• Python 3.7+ (standard library only, no external dependencies)
• Go codebase with 50+ files (100+ ideal for meaningful rules)