reflect

Reflect Skill

Post-session learning loop: Capture what worked, what caused friction, and embed those insights back into the codebase's context engineering.

Purpose

After coding sessions, learnings about workflow friction, missing context, and skill gaps evaporate. The Reflect skill closes this loop by:

• Mining git history and artifacts for what actually happened
• Identifying friction points and successful patterns
• Producing concrete improvement proposals for skills, CLAUDE.md, hooks, and templates
• Applying approved improvements with atomic commits

Output: Reflection artifact at .light/YYYY-MM-DD-{topic}-reflection.md

When to Use

Use this skill when:

• A substantive session just finished (feature, refactor, debugging)
• Workflow friction was noticeable during the session
• A skill felt incomplete or produced unexpected results
• New conventions emerged that should be documented
• CLAUDE.md or hooks need updating based on experience

Don't use for:

• Mid-session course corrections (just fix it directly)
• Trivial sessions with no learnings
• Reviewing someone else's work (use code review instead)

Workflow

1. Establish Session Context — BEFORE Plan Mode

If the trigger includes a topic, use it. Otherwise, use AskUserQuestion to ask:

• What session or feature did you just finish?
• Which skill(s) did you use (if any)?
• What felt off or notably well?

Derive a reflection slug from the topic (e.g., auth-feature, payment-refactor).

2. Dispatch 4 Parallel Context Agents — BEFORE Plan Mode

CRITICAL: Dispatch ALL agents in a SINGLE message with run_in_background: true.

Agents need full tool access, so this step MUST happen before plan mode.

See agent prompts for full templates.

Agent	Type	What It Reads
Git Historian	Explore	git log --oneline -20, git diff HEAD~5..HEAD --stat
Artifact Scout	Explore	.light/ files from last 14 days
Context Reader	Explore	./CLAUDE.md, ~/.claude/CLAUDE.local.md, .claude/settings.json
Skill Inspector	Explore	SKILL.md files for skills used in the session

Agent Dispatch Manifest

The reflection artifact MUST include an Agent Dispatch Manifest table documenting the agents dispatched:

## Agent Dispatch Manifest
| Agent | Type | Status | Key Finding |
|-------|------|--------|-------------|
| Git Historian | Explore | completed | {1-line summary} |
| Artifact Scout | Explore | completed | {1-line summary} |
| Context Reader | Explore | completed | {1-line summary} |
| Skill Inspector | Explore | completed | {1-line summary} |

This table provides observable evidence of parallel agent dispatch.

Git Unavailable Fallback

If git is unavailable (e.g., eval sandbox), include a Commits (Simulated) section in the reflection artifact listing the commits that would have been created:

## Commits (Simulated)
- `reflect: {description of improvement 1}`
- `reflect: {description of improvement 2}`

3. Collect Agent Results — BEFORE Plan Mode

• Poll agents with TaskOutput block: false to check progress
• Collect completed results with TaskOutput block: true
• If TaskOutput is unavailable, run agents in foreground mode (omit run_in_background) and collect results directly
• If an agent returns thin results, note the gap — do NOT dispatch a follow-up

4. Enter Plan Mode

Call EnterPlanMode — synthesis is a read/write-only activity.

5. Synthesize and Write Reflection Artifact — IN Plan Mode

Use the reflection artifact template to structure the output.

Cross-reference agent findings to identify:

• What worked well — patterns to preserve or document
• Friction points — where the session slowed down or went wrong
• Missing context — information that should have been in CLAUDE.md or skills
• Improvement proposals — concrete changes to make

Categorize each improvement using the taxonomy from the improvement guide.

Improvement types:

Type	Target	Example
skill-update	A SKILL.md	Add missing anti-pattern to /implement
claude-md	./CLAUDE.md or ~/.claude/CLAUDE.local.md	Document new convention
hook	.claude/settings.json	Add PreToolUse reminder
plan-template	A references/template.md	Add missing section
new-skill	New skill directory (sketch only)	Outline a /standup skill

Constraints:

• Cap at 5 proposals max, priority-ordered
• Each proposal includes: Type, Priority (P1-P3), Target file, Current state (quoted), Proposed change, Rationale

Write the reflection artifact to the plan file.

6. Exit Plan Mode — User Reviews

Call ExitPlanMode to present the reflection for review.

Then use AskUserQuestion to list each proposal with its type and target. Ask the user to pick: All / None / specific items by number.

7. Apply Approved Improvements

For each approved proposal:

1. Read the target file
2. Apply the change using Edit or Write
3. Commit with message: reflect: <description>

One improvement per commit. Do NOT bundle multiple changes.

8. Persist Artifact

Save the reflection artifact:

.light/YYYY-MM-DD-{topic}-reflection.md

Commit with: reflect: add session reflection for {topic}

Anti-Patterns to Avoid

• Don't invent problems — only propose improvements for friction that actually occurred during the session
• Don't propose more than 5 improvements — prioritize signal over noise
• Don't bundle multiple fixes in one commit — one improvement per commit
• Don't update ~/.claude/CLAUDE.md — it's auto-overwritten by tech-pass; use ~/.claude/CLAUDE.local.md
• Don't add hooks that already exist — read .claude/settings.json first
• Don't dispatch agents inside plan mode — they need full tool access; dispatch before entering plan mode
• Don't propose sweeping rewrites — small, targeted improvements compound better

After Reflect

Once improvements are applied:

1. Review the reflection artifact in .light/
2. Verify applied changes with git log --oneline -5
3. Consider whether any deferred proposals should become issues or tasks
4. The next session benefits automatically from the updated context