mrclrchtr/supi/claude-md-improver

CLAUDE.md Improver

Audit, evaluate, and improve CLAUDE.md files across a codebase to ensure PI has optimal project context.

Workflow

Phase 1: SuPi Context Baseline Review

Do not read target instruction files yet. Start from the session context and available tool surface. When code_orientation is available, assume future agents use it routinely and repeatedly—not only as a fallback. Static facts that are on demand rather than automatically injected are still poor CLAUDE.md content when Orientation exposes them directly.

Purpose: This baseline is the primary evidence for scoring Criterion 7 (SuPi Context Overlap) in Phase 3. Separate context that arrives on the first turn from facts routinely obtained through Orientation so the audit preserves human reasoning rather than generated inventories.

Step 1 — Detect SuPi context surfaces. Scan the conversation context and available tools for:

Step 2 — Build the baseline. Record both first-turn and routine on-demand coverage:

Add rows for other visible sources. Do not treat an instruction file as redundant merely because code_orientation is how its contents reach the model; compare its contents with the other facts Orientation provides.

Step 3 — Classify redundancy risk by scope. Assume heavy Orientation use:

• Root-level high risk (do NOT recommend for root ./CLAUDE.md):
• Package/module inventories, descriptions, entrypoint tables, and manifest dependency graphs
• Package paths, static project-layout trees, and direct file/directory inventories
• High-level architecture that adds no ownership, boundary, flow, exception, or rationale

• Package-specific high risk (do NOT recommend for that package's instruction file):
• Direct file/subdirectory listings
• Package manifest declarations, dependency lists, and manifest-derived relationships
• Source-symbol inventories that file Orientation exposes without additional reasoning

• Low risk (safe to recommend at any scope):
• Non-obvious commands and workflows (not routine build/test/lint)
• Gotchas, failure modes, and ordering constraints
• Ownership, boundaries, initialization/data flow, and exceptions not derivable from manifests
• Curated "start here" guidance with reasoning
• Cross-package conventions and project-specific decisions

Step 4 — Output the baseline. Produce this structured overview before Phase 2:

## Phase 1 Baseline Review
 
### SuPi Detected: yes / no
 
### SuPi Context Surfaces
 
| Source | Delivery | Already Covers | Scope |
|--------|----------|----------------|-------|
| ... | First turn / Routine Orientation | ... | Root / Package-specific |
 
### Redundancy Risk Assessment
 
-**Root-level high risk:** [categories]
-**Package-specific high risk:** [categories]
-**Low risk:** [categories]

Note: If code_orientation is available, include its documented workspace/directory/file coverage even before calling it; Phase 2 verifies the actual target outputs.

Phase 2: Discovery

Now read files from disk. Find all CLAUDE.md files in the repository:

find . \( -name "CLAUDE.md" -o -name ".claude.md" -o -name ".claude.local.md" \) \
  -not -path "*/.agents/skills/*" -not -path "*/.pi/skills/*" -not -path "*/.claude/skills/*" 2>/dev/null | head -50

Ignore instruction files inside agent skill directories. Skills may ship CLAUDE.md/AGENTS.md files as examples, templates, or fixtures. They are skill content, not project instructions — exclude them from discovery, scoring, and updates. Match only hidden agent-tool skill directories such as .agents/skills, .pi/skills, or .claude/skills; do not exclude project directories that happen to be named skills.

File Types & Locations:

Note: PI natively loads context files from the cwd and its ancestors. Nested package files are surfaced when code_orientation focuses their directories.

After discovery, run workspace code_orientation once, then directory Orientation for every target file's containing directory. Use a sufficiently high maxResults or narrower repeated focuses so a display cap is not mistaken for absence. This is required baseline evidence, not an optional removal check.

Phase 3: Quality Assessment

For each CLAUDE.md file found in Phase 2, evaluate against quality criteria, incorporating the Phase 1 baseline review results. See references/quality-criteria.md for detailed rubrics.

Quick Assessment Checklist:

Phase 1 enforcement for Criterion 7:

• Score package-specific files by comparing their contents with the non-instruction sections of directory/file Orientation. Surfacing is the delivery mechanism, not overlap by itself.
• Root- or package-level static facts classified as high risk must be flagged for removal rather than excused as minor overlap.
• Preserve concise human reasoning—ownership, boundaries, flow, exceptions, and gotchas—even when it mentions files or packages also shown by Orientation.

Quality Scores:

• A (90-100): Comprehensive, current, actionable
• B (70-89): Good coverage, minor gaps
• C (50-69): Basic info, missing key sections
• D (30-49): Sparse or outdated
• F (0-29): Missing or severely outdated

Phase 4: Quality Report Output

ALWAYS output the quality report BEFORE making any updates.

Format:

## CLAUDE.md Quality Report
 
### Summary
- Files found: X
- Average score: X/100
- Files needing update: X
- Potential token savings: ~X (removing redundant generated/discoverable context)
 
**SuPi context overlap is never "minor" when it is only a generated inventory.** Content that duplicates the first-turn overview or routine Orientation wastes context whenever the instruction file is loaded and MUST be flagged for removal.
 
### File-by-File Assessment
 
#### 1. ./CLAUDE.md (Project Root)
**Score: XX/100 (Grade: X)**
 
**Context Overlap Review:**
- **Fully redundant (root-level):** [sections already covered by baseline context — applies to root `./CLAUDE.md`]
- **Fully redundant (package-specific):** [sections already covered by baseline context — applies to that package's `CLAUDE.md`]
- **Partially redundant:** [sections with overlap plus human-only value]
- **Unique:** [sections that should stay]
- **Estimated waste:** ~X tokens (characters ÷ 4) duplicate first-turn or routine Orientation context — should be removed
 
| Criterion | Score | Notes |
|-----------|-------|-------|
| Commands/workflows | X/15 | ... |
| Architecture clarity | X/15 | ... |
| Non-obvious patterns | X/15 | ... |
| Conciseness | X/15 | ... |
| Currency | X/15 | ... |
| Actionability | X/15 | ... |
| SuPi context overlap | X/10 | ... |
 
**Issues:**
- [List specific problems]
 
**Recommended removals (non-negotiable generated/discoverable overlap):**
- [List static inventories that MUST be removed or compressed, with estimated token savings.]
 
**Recommended additions:**
- [List what should be added]
 
#### 2. ./packages/api/CLAUDE.md (Package-specific)
...

Phase 5: Targeted Updates

Core principle: every token must earn its place in the instruction file. If content doesn't save future sessions more time than it costs to read, remove it. No instruction file should exceed 200 lines — above that, every line must fight for its place against removal.

After outputting the quality report, ask user for confirmation before updating.

Before recommending removals, verify the Phase 2 Orientation baseline. Compare each target file with workspace and focused code_orientation output. Remove facts Orientation already supplies; retain human-only meaning. Never count the instruction text shown inside the Orientation result as evidence that the same text is redundant.

Update Guidelines (Critical):

1. Remove or compress unnecessary content first — Before adding anything, flag sections that MUST be removed or tightened. Never skip removals because of edit churn—a one-time edit that saves tokens on every relevant load pays for itself quickly.

• Routine command listings (npm install, npm test, npm run build) — remove; they're in package.json
• Package/module inventories that duplicate the complete first-turn overview — MUST be removed
• Package paths, project trees, manifest facts, and dependency listings available through routine code_orientation use — MUST be removed unless they add ownership, flow, boundary, exception, or rationale
• Verbose explanations where a one-liner suffices — compress
• Stale or outdated commands, file references, or architecture descriptions — remove