| name | improve-skills |
| description | Improve existing skills based on the current session. Use at the end of a session (or when the user asks) to capture new debugging patterns, data issues, data validation tracks, querying techniques, doc references, or workflow improvements learned during the session. Keeps skills lean and personalized. |
Improve skills from session
Review the current session and update skills with essential learnings. Skills are the team's institutional memory — keep them lean, specific, and actionable.
Principles
- Read the skill first — always read the current SKILL.md before proposing changes. Don't duplicate what's already there.
- Lean updates only — add the minimum needed. A single bullet point or a 3-line code snippet is often enough.
- Real problems only — only add things that actually happened in this session or that the user explicitly asked to capture. No hypothetical scenarios.
- Doc references matter — if you found a docs page that was essential to solving a problem, add it as an "Essential Reading" link.
- Tools and commands — if a specific CLI command or MCP tool was key to diagnosing an issue, add it.
- Don't restructure — don't reorganize or rewrite existing skill content. Append to the right section or add a new subsection.
Process
1. Scan the session for learnings
First identify which skills are active in the current toolkit — check the installed skills directory or use the toolkit_info MCP tool. Then review the conversation for:
Errors and debugging:
- New error types, root causes, and the commands or MCP tools that diagnosed them
- Config settings that helped (e.g., verbosity, timeouts, flags)
- Workarounds for API, source, or destination-specific behaviors
Data and schema:
- Unexpected data types or coercions needed
- Schema surprises (nesting, missing columns, naming conventions)
- Processing patterns that worked
Data access and querying:
- Library-specific gotchas (e.g., ibis, dlt dataset API)
- Useful query patterns or MCP tool calls
Source and pipeline configuration:
- Auth, pagination, or rate-limit quirks
- Config resolution surprises
- Source/resource parameterization patterns
Workflow:
- Missing cross-references between skills
- New skills that should be in the workflow
- Steps that are in the wrong order
Map each learning to the most relevant skill in the active toolkit.
2. Read the target skills
For each learning, read the relevant SKILL.md. Check:
- Is this already covered?
- Where does it fit best? (which section)
- Would it contradict or duplicate existing content?
3. Propose changes to the user
Present a summary of proposed updates:
Proposed skill updates:
<skill-name>:
+ [section] Added: <brief description>
<skill-name>:
(no changes)
<skill-name>:
+ [section] Added: <brief description>
Get user approval before editing. The user may want to adjust, skip, or add more.
4. Apply changes
Edit the SKILL.md files. For each change:
- Add to the most relevant existing section
- If no section fits, add a new subsection at the end (before "Next steps" if present)
- Include doc links as
**Ref:** <url> or **Essential Reading:** <url>
- Keep code examples minimal — 3-5 lines max
5. Update workflow if needed
Check if workflow.md cross-references are still accurate after skill changes. Update if new skills were added or handoff points changed.