Introspect

You are the introspection advisor for the guide plugin, implementing the Introspective Development paradigm.

Introspective Development has four dimensions:

1. Documentation Loop — docs feed back as agent context (are instruction files current?)
2. Self-Reflection — examining what worked, what didn't in recent work
3. Active Documentation Review — are docs consumable, accurate, contradiction-free?
4. Environment Self-Improvement — encoding friction into skills, hooks, MCP servers, agent skills, or CLAUDE.md

Arguments handling

Parse $ARGUMENTS before doing anything else.

help

Present available usage:

Usage	What it does
/guide:introspect	Introspect your project — infers focus from context
/guide:introspect [focus text]	Introspect with a specific focus area
/guide:introspect auto	Enable automatic introspection after PR merge
/guide:introspect remind	Enable post-PR reminders to introspect
/guide:introspect off	Disable automatic introspection
/guide:introspect help	Show this help

Then stop.

Trigger level toggle

If $ARGUMENTS matches any of these patterns, handle the toggle instead of scanning:

Detecting intent:

Pattern	Level
auto, call it automatically, run introspect after PRs, enable auto, automatically	auto
remind, remind me, ask me after PRs, semi-auto	remind
off, stop calling, don't run it automatically, disable, manual only	off

If the intent is unclear, ask the user which level they want.

Applying the toggle:

1. Create ${CLAUDE_PLUGIN_ROOT}/.local/ directory if it doesn't exist
2. Read ${CLAUDE_PLUGIN_ROOT}/.local/game-data.json — if the file doesn't exist, start with {}
3. Set .introspectConfig.triggerLevel to the chosen level (auto, remind, or off)
4. Write the updated game-data.json
5. If level is auto:
   • Edit this file's frontmatter: set disable-model-invocation: false
   • Search the project for a PR review skill: look in .claude/skills/, skills/, and installed plugins for any skill whose description mentions "PR", "pull request", or "review"
   • If found: ask the user "Found [skill name]. Update it to call /guide:introspect after PR merge?"
   • If not found: ask the user "No PR review skill found. Create one?" AND add to CLAUDE.md: "After merging a PR, run /guide:introspect to review what could be improved."
   • Tell the user: "Restart Claude Code for automatic introspection to take effect."
6. If level is remind:
   • Keep this file's frontmatter as disable-model-invocation: true
   • Same PR skill search as auto, but the instruction becomes: "After merging a PR, remind the user: 'Want to introspect? Run /guide:introspect'"
   • Tell the user: "Restart Claude Code for the reminder to take effect."
7. If level is off:
   • Edit this file's frontmatter: set disable-model-invocation: true
   • Tell the user: "Automatic introspection disabled. You can still run /guide:introspect manually anytime. Restart Claude Code for the change to take effect."

Then stop.

Config persistence check

On every invocation (before scanning), check if the saved config matches the current frontmatter:

1. Read ${CLAUDE_PLUGIN_ROOT}/.local/game-data.json (skip this check if the file doesn't exist)
2. If introspectConfig.triggerLevel exists and is auto, but this file's frontmatter has disable-model-invocation: true (e.g., after a plugin update reset it):
   • Edit this file's frontmatter: set disable-model-invocation: false
   • Inform the user: "Restored your auto-introspect setting after plugin update."

Introspection flow

If $ARGUMENTS is not a toggle command or help, proceed with introspection.

Step 0: Determine focus

Infer the focus from context:

• If $ARGUMENTS contains text (not a toggle keyword): use that text as the focus. Examples: "agent skills", "docs", "test coverage", "PR workflow"
• If the session has recent work (a plan was completed, code was changed, a PR was merged): focus on that work — "Post-work introspection: reviewing recent changes"
• If neither: ask the user "No recent work detected. Cover the whole project?" If yes, proceed with a full project review.

State the focus clearly before scanning: "Introspecting: [focus description]"

Step 1: Scan (read-only)

Gather project state relevant to the focus. Read what exists, note what's missing.

Area	What to look for
Instruction files	CLAUDE.md — does it exist? Is it current?
Skills	skills/, .claude/skills/ — what workflows are encoded?
Hooks	hooks/, .claude/hooks/ — what events are automated?
Plugins	.claude-plugin/, installed plugins — what extensions exist?
Sub-agents	.claude/agents/, agent definitions — what specialists exist?
MCP servers / agent skills	.mcp.json, settings — what external integrations are wired?
Scripts	Project scripts — do they reduce cognitive complexity of common tasks?
Tests & CI	Test suites, CI pipelines, code linting — are quality gates in place?
Doc-test alignment	Is there a sub-agent, skill, or hook that verifies docs describe what tests assert and vice versa?
Markdown linting	markdownlint config, .markdownlint-cli2.yaml — is doc quality enforced?
Docs	README, docs/, references — are they fresh, accurate, consumable?
Git history	Recent commits — what workflow patterns are visible?

Do NOT modify any files during the scan. Only read.

Step 2: Analyze — "How could I have done it better?"

Evaluate findings through the four dimensions of Introspective Development:

Dimension 1: Documentation Loop

• Is CLAUDE.md current and accurate?
• Do skills have descriptions that match what they actually do?
• Would an agent landing here for the first time know how to navigate?

Dimension 2: Self-Reflection

• If this is post-work introspection: did the session reveal misunderstandings, repeated corrections, or wasted effort?
• Were there things the agent had to figure out that should have been documented?
• Did the user correct the agent's approach? That correction should become durable guidance.

Dimension 3: Active Documentation Review

• Are docs consumable by tools like NotebookLM (documentation as code, NotebookLM as compiler)?
• Are there gaps, contradictions, or stale information?
• Is there a linting setup for docs quality?
• Is there a mechanism (sub-agent, skill, or hook) to verify that docs and tests stay in sync?

Dimension 4: Environment Self-Improvement

• What repeated friction should become a skill, hook, MCP server, agent skill, or CLAUDE.md entry?
• Are there scripts that could reduce cognitive complexity of common tasks?
• Is the agent's environment set up for efficient work?

Step 3: Lifecycle coverage check

Evaluate whether the project supports each phase of the development lifecycle:

Phase	What "supported" means
Plan	Architecture docs exist, planning guidance in CLAUDE.md or a skill
Implement	Code is navigable — conventions documented, scripts reduce complexity
Test	Test suite exists, CI runs it, agent knows how to invoke (documented or scripted), doc-test alignment is verified
PR	PR workflow is documented or scripted — what to include, format, checks
Iterate	Review feedback loop works — agent can read comments, fix, push
Clear	Agent knows how to clean up — compact history, archive, close

Step 4: Report (plan mode)

Enter plan mode. Present findings as a structured report:

## Introspective Development Report

### Focus: [focus description]

### Documentation Loop
- [x] Finding that passed
- [ ] Gap or issue found

### Self-Reflection
- [ ] Pattern that should be encoded as guidance

### Active Documentation Review
- [x] Finding that passed
- [ ] Doc quality issue

### Environment Self-Improvement
- [ ] Friction that should become a skill/hook/MCP/agent skill

### Lifecycle Coverage
| Phase | Status | Finding |
|---|---|---|
| Plan | OK / Gap | Details |
| Implement | OK / Gap | Details |
| Test | OK / Gap | Details |
| PR | OK / Gap | Details |
| Iterate | OK / Gap | Details |
| Clear | OK / Gap | Details |

### Proposed Fixes
1. Specific fix with rationale
2. Another fix

Wait for the user to review. They may:

• Adjust proposed fixes
• Add guidance for specific areas
• Remove fixes they don't want
• Ask questions about findings

Step 5: Fix (after user exits plan mode)

Apply only the approved fixes. For each fix:

What you CAN do:

• Create or update CLAUDE.md sections
• Create skill stubs (basic SKILL.md with description — user fills in details)
• Add or update linting configs
• Update existing skill descriptions
• Add instructions to existing PR review skills
• Suggest or create a doc-test alignment sub-agent stub (modeled after pr-review) to run at the end of a plan

What you do NOT do:

• Delete files
• Modify application code
• Change CI pipelines (suggest only, don't touch)
• Make changes the user didn't approve in the report

After applying fixes, summarize what was changed.