reflect

name	reflect
description	Use when auditing, improving, or maintaining project health — CLAUDE.md accuracy, memory hygiene, settings, plans, and pattern discovery. Invoke periodically or when the user asks to audit or improve their project configuration.

Reflect

Project self-improvement: audit configs against reality, discover patterns worth documenting, clean up stale artifacts.

Input

The user input is: $ARGUMENTS

Mode Detection

• Empty or full: Run all phases
• health: Phase 1 only (project health audit)
• evolve: Phase 2 only (pattern discovery + CLAUDE.md evolution)
• cleanup: Phase 3 only (housekeeping)

Phase 1: Project Health

Audit the current project's Claude configuration against reality.

1.1 CLAUDE.md Accuracy

Read the project's CLAUDE.md and verify against the actual codebase:

• Module list: Does the listed module/domain structure match what actually exists?
• File paths: Do referenced paths (shared components, config files) actually exist?
• Patterns: Do described patterns (query factories, service layer, form fields) match current implementations?
• Shared components/hooks: Are listed components and hooks still accurate?
• Outdated rules: Are there rules referencing libraries, tools, or conventions no longer in use?

1.2 Parent vs Project CLAUDE.md

Check content is at the correct level:

• Is there generic content in the project CLAUDE.md that belongs in the parent?
• Is there project-specific content in the parent CLAUDE.md?
• Are there rules duplicated between both files?

1.3 Memory Hygiene

Read all memory files in the project's memory directory:

• Flag stale entries (reference removed features, old patterns)
• Flag duplicates
• Flag entries that conflict with current codebase state

1.4 Settings Audit

Read settings.json and settings.local.json:

• Are all referenced plugins still installed and relevant?
• Are there permissions for tools/commands no longer used?

1.5 Plans & Prompts

For each plan in .claude/plans/:

1. Check if the feature has been implemented (search codebase)
2. Categorize: Completed, Partially completed, Not started, Stale
3. Propose archiving completed plans to .claude/plans/archive/

For each prompt in .claude/prompts/:

1. Check if the feature has been implemented
2. If implemented, propose archiving

1.6 Project Contracts Freshness

Audit the one-time, project-level contract files that downstream craft-skills consume. Each has the same lifecycle: generated once per project, hand-editable, needs occasional refresh when the underlying reality changes.

.claude/aesthetic-direction.md:

• If missing: suggest running craft-skills:aesthetic-direction once (only relevant for projects with a UI).
• If present and older than 180 days: ask whether the design system has changed since; offer regeneration.

.claude/reuse-index.md:

• If missing: suggest running craft-skills:reuse-index once. Without it, every planning session falls back to ad-hoc grep.
• If present and older than 60 days: compare entries against the current shared directory. If >5 top-level shared files exist whose exports aren't in the index, flag it as stale and suggest craft-skills:reuse-index --force.
• If the index's "What NOT to duplicate" section has hand-edited entries, preserve them across regeneration — call this out when suggesting a refresh.

CLAUDE.md (parent + project): already covered by 1.1 and 1.2 above.

Freshness is advisory, not enforced — reflect only suggests, the human decides.

Phase 2: Evolve

Discover patterns in how you work and propose improvements to project configuration.

2.0 Upstream Sync Check

craft-skills absorbs methodology from the upstream superpowers plugin. Before discovering local patterns, check whether a newer upstream version is available:

CRAFT_SCRIPTS=$(find ~/.claude/plugins -name "sync-check.sh" -path "*/craft-skills/*" -exec dirname {} \; 2>/dev/null | head -1)
[ -n "$CRAFT_SCRIPTS" ] && bash "$CRAFT_SCRIPTS/sync-check.sh" || echo "sync-check unavailable — skip"

If it reports a new version, note it: absorbing upstream changes is tracked in the plugin's references/superpowers-sync.md (see the changelog there for what's already been adopted).

2.1 Pattern Discovery

Dispatch an agent to perform analysis of the current project:

1. Read all memory files in the project's memory directory
2. Analyze recent git history (last 2 weeks)
3. Identify patterns:
   • Repeated feedback → should become a rule in CLAUDE.md
   • Recurring manual tasks → worth documenting as a workflow in CLAUDE.md
   • Common architectural patterns emerging → should be documented in CLAUDE.md
   • Frequent mistakes or corrections → add guard rules to CLAUDE.md
4. Generate proposals ranked by impact
5. Present to user for approval

2.2 Cross-Project Insights

If the project has a parent CLAUDE.md (monorepo or workspace):

1. Read memory files from sibling projects under the same parent directory
2. Look for patterns that repeat across projects:
   • Same feedback given in multiple projects → belongs in parent CLAUDE.md
   • Same workarounds applied → should be a shared rule
   • Conventions that emerged independently → formalize in parent CLAUDE.md
3. Propose updates to the parent CLAUDE.md (never to sibling project files)

2.3 CLAUDE.md Completeness

Compare what's documented vs what the codebase actually does:

• Are there undocumented conventions the code follows consistently?
• Are there utility functions/hooks widely used but not mentioned?
• Are there testing patterns or data-fetching patterns worth documenting?

Propose additions to CLAUDE.md for anything that would help Claude work more effectively in this project.

Phase 3: Cleanup

3.1 Auto-Fix (low risk, verifiable)

Applied automatically:

• Update domain/module lists in CLAUDE.md to match actual directory structure
• Fix stale file paths in CLAUDE.md
• Archive completed plans to .claude/plans/archive/

3.2 Ask First (subjective or significant)

Presented for approval:

• Removing stale memory entries
• Moving content between parent/project CLAUDE.md
• Adding new rules or sections to CLAUDE.md
• Removing outdated rules from CLAUDE.md

Presenting Findings

For each issue found, present:

• What's wrong: The specific inconsistency
• Where: Which file(s) are affected
• Proposed fix: The concrete change
• Impact: How this affects Claude's behavior

Categorize as:

• Critical — causes errors, broken paths, missing principles
• Incorrect — factually wrong, will cause confusion
• Stale — outdated references
• Duplication — rules repeated across levels
• Cleanup — plans/prompts/memory that can be archived
• Improvement — opportunities to enhance CLAUDE.md effectiveness

Wait for user approval before making changes (except auto-fix tier).