lesson-gen

name	lesson-gen
description	Mid-session quick capture — save an insight before the context moves on
version	0.2.0
level	1
triggers	["extract pattern","save this approach","log this","note this"]
context_files	["context/learnings.md"]
steps	[{"name":"Identify","description":"What specifically happened? State it in one sentence."},{"name":"Quality Gate","description":"Apply the 3-point test — can't Google it, codebase-specific, took real effort"},{"name":"Generalize","description":"State the general principle, not just the session-specific instance"},{"name":"Write","description":"Append to context/learnings.md in standard format"}]

Lesson-Gen Skill

Mid-session quick capture. Use when something is worth keeping but you do not want to break flow to do a full wrap-up.

The 3-Point Quality Gate

1. Could someone Google this in 5 minutes? → Must be NO
2. Is this specific to this project or codebase? → Must be YES
3. Did this take real effort to discover? → Must be YES

4-Verdict System

After applying the quality gate, classify the learning into one of four verdicts:

1. SAVE

Passes all 3 quality gate points. Save immediately to context/learnings.md in standard format.

2. IMPROVE THEN SAVE

Good insight but needs refinement:

• Statement is too vague or session-specific
• Missing the general principle
• Lacks concrete details (file paths, numbers, specific errors)

Action: Rewrite to be more concrete/general, then save to context/learnings.md.

Example:

• Before: "Database queries were slow"
• After: "PostgreSQL queries without indexes on user_id field caused 500ms+ latency on /api/users endpoint. Added index, reduced to 50ms."

3. ABSORB INTO EXISTING SKILL

This learning belongs in a skill, not in context/learnings.md:

• It's a reusable technique or pattern
• It applies across projects, not just this one
• It's procedural (how to do X) rather than factual (X happened)

Action: Identify which existing skill this belongs in (or propose a new skill if none fit). Add to that skill's body or examples section.

Example:

• Learning: "When debugging API issues, always check network tab first, then server logs, then database queries"
• Verdict: ABSORB INTO debug-session skill (this is a general debugging procedure)

4. DROP

Fails the quality gate:

• Can be Googled easily
• Too generic (not codebase-specific)
• Low effort / obvious

Action: Discard. Do not write to any file.

Entry Format

[YYYY-MM-DD] [category] — [specific statement]

One entry per insight. Do not bundle.

Anti-Patterns

Do not write the session-specific instance — write the general principle it illustrates.

Do not write what you intended to do — write what actually happened and what it revealed.

Mandatory Checklist

1. Verify the learning was evaluated against all 3 quality gate points
2. Verify one of the 4 verdicts was chosen (SAVE / IMPROVE THEN SAVE / ABSORB / DROP)
3. Verify SAVE verdicts were appended to context/learnings.md in correct format
4. Verify IMPROVE THEN SAVE verdicts were refined before saving
5. Verify ABSORB verdicts identified the target skill
6. Verify DROP verdicts were not written anywhere