When to use: You've told Claude 'use lodash isEmpty, not !arr.length' three times and it keeps reverting.
Flow
-
Capture the next correction
Use pro-workflow. Capture this rule: 'Use lodash isEmpty for emptiness checks; never !arr.length'. Scope: src/utils/.✓ Copied
→ Rule saved with scope; CLAUDE.md updated
-
Verify next session
Tomorrow: ask Claude to clean up an empty-check pattern in src/utils/.✓ Copied
→ Claude uses isEmpty without re-prompting
-
Audit the rule set monthly
List all captured rules. Anything stale or superseded?✓ Copied
→ Rule list with last-applied dates; prune cruft
Outcome: Each correction is paid once, not weekly.
Pitfalls
- Over-broad rules ('always use functional style') backfire — Scope rules to directories or feature areas; concrete > general
When to use: You want to refactor module A while Claude continues a feature on module B.
Flow
-
Spawn worktrees
Use pro-workflow. Create two worktrees: ../wt-refactor-a and ../wt-feature-b from main.✓ Copied
→ Worktrees created; rules synced into each
-
Run threads
Start a Claude Code session in each. Worktree-aware rules apply per directory.✓ Copied
→ Both sessions running; no merge conflicts mid-flight
-
Reintegrate
When both branches are ready, open PRs. The skill notes any rules that triggered and their effects.✓ Copied
→ PRs include rule-effect summaries
Outcome: Real parallelism — agent throughput multiplied without coordination tax.
Pitfalls
- Rules drift between worktrees if only updated in one — The skill syncs rule set on worktree creation; for ongoing edits, run sync-rules
When to use: You want data on whether 'AI gets smarter on this repo over time' is real or vibes.
Flow
-
Baseline
Use pro-workflow. Show me corrections per session over the last 30 days.✓ Copied
→ Trend chart; expect downward slope
-
Tag categories
Categorize corrections: style, architecture, dependency choice, test pattern. Show category-wise decay.✓ Copied
→ Per-category trend
-
Decide
Recommend whether to roll this out team-wide based on data.✓ Copied
→ Concrete recommendation with caveats
Outcome: Data-grounded decision on agent workflow investment.
Pitfalls
- Drop in corrections != quality up — could just mean you stopped reviewing — Cross-check against PR review feedback or test failure rate