| name | self-improve |
| description | Introspect on a completed work session to propose and apply improvements to documentation, agent guidance, and repo quality |
| outputs | [{"name":"improvements","description":"List of proposed changes with rationale, and which were applied"}] |
Purpose
Guide end-of-session reflection that turns work context into concrete improvements to documentation and agent guidance. Instead of maintaining a running session log, this skill reconstructs what happened from git state and conversation context, then proposes targeted edits.
Preconditions
- At least one commit exists on the current branch beyond main (or there are uncommitted changes)
- The agent has context from the current conversation about work performed
Steps
1. Gather session context
Reconstruct what happened during the session using git state:
git diff main...HEAD --stat
git log main..HEAD --oneline
git diff --stat
git diff --cached --stat
Read the changed files to understand the nature of the work.
Build a mental model of:
| Element | Source |
|---|
| What code was added/changed | git diff |
| Why it was changed | Commit messages + conversation context |
| What patterns were used | Reading the changed files |
| What friction was encountered | Conversation context (errors, retries, pivots) |
2. Audit documentation freshness
For each file changed, check whether corresponding documentation is stale:
| If this changed... | Check these docs... |
|---|
Routes or services (src/routes/, src/services/) | docs/openapi/ for API spec accuracy, AGENTS.md architecture section |
Models or migrations (src/models/, src/migrations/) | docs/tech-stack.md, AGENTS.md database section |
| Test patterns or helpers | docs/guides/testing.md, best_practices.md testing section |
| Infrastructure or config (Dockerfile, CI) | docs/guides/infrastructure.md, docs/guides/dev-setup.md |
| Frontend components or pages | AGENTS.md frontend section |
| Worker or queue changes | AGENTS.md worker section |
| New dependencies (package.json) | docs/tech-stack.md |
| Dev workflow (scripts, Makefile, docker-compose) | docs/guides/yarn-commands.md, docs/guides/dev-setup.md |
Read the relevant doc files and compare their descriptions against the current code state.
3. Check for common doc debt signals
Scan documentation files for known debt patterns
Stale references:
- File paths mentioned in docs that no longer exist
- Command examples that reference renamed scripts
- Architecture descriptions that don't match current directory structure
Duplicated content:
- Check for substantially similar paragraphs across
AGENTS.md, README.md, and doc files
- Identify content that should live in one place and be referenced from others
Unresolved markers:
grep -rn 'TODO\|TBD\|FIXME\|HACK\|XXX' docs/ AGENTS.md best_practices.md
Check if any of these can now be resolved given the session's work.
4. Review agent guidance gaps
Reflect on the session's conversation context for friction signals:
| Friction signal | Potential improvement |
|---|
| Agent had to search extensively for a pattern | Add the pattern to AGENTS.md or best_practices.md |
| Agent made incorrect assumptions about architecture | Clarify the relevant AGENTS.md section |
| Agent used a workaround for a known issue | Document the gotcha in the relevant guide |
| A new convention was established during the session | Capture it in best_practices.md |
| Agent needed information that wasn't in any doc | Add it to the appropriate file |
| Agent had to ask for clarification about project norms | Document the answer for future sessions |
| Testing required unexpected setup or teardown | Add to docs/guides/testing.md |
5. Propose improvements
Present a numbered list of proposed changes. Each proposal should include:
### Proposal N: [Brief title]
- **Target file:** `path/to/file.md`
- **Category:** [docs-freshness | broken-link | duplicate-content | new-pattern | agent-guidance | stale-reference]
- **What:** [Specific change to make]
- **Why:** [Rationale tied to session context or audit finding]
Categories explained:
| Category | Meaning |
|---|
docs-freshness | Documentation doesn't match current code state |
broken-link | A relative link points to a file that doesn't exist |
duplicate-content | Same information repeated across multiple files |
new-pattern | A pattern or convention established during the session worth capturing |
agent-guidance | Missing or misleading guidance in agent-facing files |
stale-reference | Doc references a file, command, or concept that no longer exists |
If no improvements are found, state "No improvements identified" with a brief explanation of what was checked.
6. Apply approved changes
Ask the user which proposals to apply:
- "Apply all" — edit all target files
- Selective — apply only chosen proposals
- "None" — skip application, keep proposals as a record
For each approved proposal, edit the target file directly. Keep changes minimal and focused — don't rewrite entire sections when a sentence fix suffices.
7. Summary
Output a final summary:
## Session Reflection Summary
### Applied
- [List of proposals that were applied, with file paths]
### Deferred
- [List of proposals not applied, with brief reason]
### Session stats
- Files changed this session: N
- Docs checked: N
- Proposals generated: N
- Proposals applied: N
Completion Criteria
