Continuous Improvement

Continuous Improvement Skill

Orchestrates periodic quality improvement cycles. Combines quality audits, harness tuning, competitive research, and process refinement into a unified improvement loop. Use this skill to keep the system getting better over time across all dimensions.

When to Use

• After every 5 versions (v0.0.15, v0.0.20, v0.0.25...)
• When agent performance degrades noticeably
• When starting a new sprint or major feature
• When the user asks to "make things better" or "improve quality"

The Improvement Cycle

┌──────────────────────────────────────────────────────────┐
│                                                          │
│   1. MEASURE         2. ANALYZE         3. IMPROVE       │
│   ─────────         ─────────          ─────────         │
│   Quality Audit     Root Cause          Fix Top 3        │
│   Harness Audit     Prioritize          Update Config    │
│   Metrics Gather    Compare to Last     Create Issues    │
│                                                          │
│                     4. VERIFY                             │
│                     ─────────                             │
│                     Re-run Audit                          │
│                     Scores Improved?                      │
│                     Commit Changes                        │
│                                                          │
└──────────────────────────────────────────────────────────┘

Phase 1: MEASURE

Run all three audits and collect metrics:

1a. Quality Audit (use quality-audit skill)

• Engineering Robustness score (0-10)
• UI/UX Aesthetics score (0-10)
• Feature Completeness score (0-10)

1b. Harness Engineering Audit (use harness-engineering skill)

• CLAUDE.md size (target ≤120 lines)
• Hook coverage
• State file freshness
• Agent performance metrics

1c. Metrics Collection

# Test health
npm test 2>&1 | tail -5                              # Pass/fail count
npx tsc --noEmit 2>&1 | grep -c "error TS" || echo 0 # Type errors

# Codebase health (use Glob and Grep tools, not bash find/grep)
# Use: Glob("src/**/*.{ts,tsx}") to count files
# Use: Grep("pattern", path) to search content

# Agent health (if git log available)
git log --oneline --since="2 weeks ago" | wc -l        # Recent commits
git log --oneline --since="2 weeks ago" --grep="fix:" | wc -l  # Fix ratio

Phase 2: ANALYZE

Compare to Previous Audit

Read .llm/reports/ for the last quality audit report. Compare scores:

• Which dimensions improved?
• Which regressed?
• What changed since last audit?

Root Cause Analysis

For each dimension that scored < 7 or regressed:

1. List the specific checklist items that failed
2. Trace each failure to a root cause:
   • Process gap: No hook/CI check for this
   • Knowledge gap: Agent didn't know the standard
   • Complexity gap: Task was too large for one session
   • Tooling gap: No automated way to check this

Prioritize

Rank improvements by: (impact on score) × (ease of fix) / (risk of regression)

Phase 3: IMPROVE

Fix Top 3 Issues

For each of the top 3 prioritized improvements:

1. If process gap: Add a hook, CI check, or agent rule
2. If knowledge gap: Add to reference doc or agent definition
3. If complexity gap: Break into smaller tasks, add subagent
4. If tooling gap: Create a script or skill

Update Configuration

• Update .claude/settings.json hooks if new enforcement needed
• Update agent definitions if new guidance needed
• Update reference docs if standards changed
• Create GitHub issues for remaining items

Record Changes

Write improvement report to .llm/reports/improvement-cycle-YYYY-MM-DD.md:

# Improvement Cycle — [date]

## Previous Scores → Current Scores
| Dimension | Previous | Current | Delta |
|-----------|----------|---------|-------|
| Engineering | X | Y | +/-Z |
| UI/UX | X | Y | +/-Z |
| Completeness | X | Y | +/-Z |

## Changes Made
1. [change] — addresses [root cause]

## Issues Created
- #NNN: [title]

## Next Cycle Focus
- [what to prioritize next time]

Phase 4: VERIFY

• Re-run the specific checks related to changes made
• Confirm scores improved (or at least didn't regress)
• Commit all configuration changes
• Push and create PR if appropriate

Automation: Scheduled Improvement Cycle

To run this automatically, configure a scheduled trigger:

{
  "name": "improvement-cycle",
  "schedule": "0 9 * * 5",
  "prompt": "Run the continuous-improvement skill: measure all three dimensions, analyze against last report, fix top 3 issues, and verify improvements.",
  "description": "Weekly quality improvement on Friday mornings"
}

Quality Score History

Maintain a running log in .llm/reports/quality-scores.csv:

date,engineering,ux,completeness,overall,notes
2026-03-29,7,6,5,6,baseline audit

This enables trend tracking across versions.