jurislm learn-eval — Extract, Evaluate, then Save
Analyze the current session for reusable patterns, apply a quality gate, determine the right save location (Global vs Project), then write the skill file after user confirmation.
What to Extract
Scan the session for:
- Error Resolution Patterns — root cause + fix + why it's reusable
- Debugging Techniques — non-obvious diagnostic steps, tool combinations
- Workarounds — library quirks, API limitations, version-specific fixes
- Project-Specific Patterns — conventions, architecture decisions, integration patterns
Do not extract: typos, one-off syntax errors, or issues caused by a specific outage.
Process
Step 1: Review the Session
Read the session conversation and identify the most valuable/reusable insight. Focus on insights that would save time in a future session, not on what was done.
Step 2: Draft the Skill File
Use this format:
---
name: pattern-name
description: "Under 130 characters — what this pattern solves"
user-invocable: false
origin: auto-extracted
---
# [Descriptive Pattern Name]
**Extracted:** [Date]
**Context:** [When this applies]
## Problem
[Specific problem this solves]
## Solution
[Pattern/technique/workaround — with code examples]
## When to Use
[Trigger conditions for future sessions]
Step 3: Quality Gate
Execute all checklist items before proceeding:
-
grep ~/.claude/skills/and~/.claude/rules/and relevant project.claude/skills/and.claude/rules/by keyword — check for content overlap - Check
MEMORY.md(both project and global) for overlap - Consider whether appending to an existing skill would suffice
- Confirm this is a reusable pattern, not a one-off fix
Then issue one holistic verdict:
| Verdict | Meaning | Next Action |
|---|---|---|
| Save | Unique, specific, well-scoped | Proceed to Step 4 |
| Improve then Save | Valuable but needs refinement | List improvements → revise → re-evaluate once |
| Absorb into [X] | Should append to an existing skill | Show target + additions → Step 4 |
| Drop | Trivial, redundant, or too abstract | Explain reasoning and stop |
Step 4: Determine Save Location
Ask: "Would this pattern be useful in a different project?"
| Answer | Location |
|---|---|
| Yes — generic (bash, LLM API behavior, debugging technique) | ~/.claude/skills/learned/ (Global) |
| No — project-specific (config quirks, project architecture) | .claude/skills/learned/ (Project) |
When unsure, choose Global. Moving Global → Project is easier than the reverse.
Step 5: Present and Confirm
Output format before saving:
### Checklist
- [x] skills/ grep: no overlap (or: overlap found → details)
- [x] MEMORY.md: no overlap (or: overlap found → details)
- [x] Existing skill append: new file appropriate (or: should append to [X])
- [x] Reusability: confirmed (or: one-off → Drop)
### Verdict: Save / Improve then Save / Absorb into [X] / Drop
**Rationale:** (1-2 sentences)
**Save path:** ~/.claude/skills/learned/pattern-name.md
**Draft:**
[full skill file content]
Save only after explicit user confirmation. Do not write files speculatively.
Step 6: Save
Write to the determined path. For Absorb, append to the existing skill file rather than creating a new one.
Design Notes
- Keep skills focused — one pattern per file
- When verdict is Absorb, prefer appending with a clear separator (
---) and heading - The quality gate uses holistic judgment, not numeric scores — the checklist ensures no step is skipped