learn

You are capturing session learnings for ConvoPilot.

Action: $ARGUMENTS

---

Two modes

• /learn log — Scan the current session, find learnings, save to working_docs/session-log.md
• /learn apply — Review saved learnings and write the important ones into CLAUDE.md

---

/learn log — Capture from this session

Step 1 — Scan the session for these learning types

Type	What to look for
DECISION	A choice was made (e.g. "we're not using OpenAI, only Claude + Gemini")
MISTAKE	Something was done wrong and corrected (e.g. "used the wrong Langfuse host for traces")
PATTERN	A reusable approach discovered (e.g. "use @trace_root/@trace_span decorators for tracing")
CONVENTION	A rule that should be followed going forward (e.g. "always check existing services for reusable patterns first")

Step 2 — For each learning found, ask the owner

"I found this learning: [description]. Should I log it? (yes / skip)"

Step 3 — Append approved learnings to working_docs/session-log.md

Create the file if it doesn't exist. Format:

### L-{N}: {short title}
- **Type:** DECISION | MISTAKE | PATTERN | CONVENTION
- **Date:** {today}
- **What happened:** {1–2 sentences}
- **Where to apply:** CLAUDE.md | skill file | both
- **Proposed rule:** {the actual instruction to add}
- **Status:** captured

Step 4 — Summary

Tell the owner: "Logged {N} learnings to working_docs/session-log.md. Run /learn apply to write them into CLAUDE.md."

---

/learn apply — Write learnings to CLAUDE.md

Step 1 — Read working_docs/session-log.md

Extract all entries with Status: captured.

Step 2 — Present a table to the owner (GATE — wait for approval)

#	Learning	Target	Draft rule
L-1	No OpenAI key	CLAUDE.md	"ConvoPilot uses Anthropic Claude and Google Gemini only. No OpenAI."
L-2	Langfuse stage TBD	CLAUDE.md	"Langfuse keys are TBD — do not use prod keys. Await stage project."

Ask: "Should I apply these to CLAUDE.md? (confirm each or say 'all')"

Step 3 — Apply approved learnings

Read the target file, add the rule in the appropriate section, write back.

Rules for where things go in CLAUDE.md:

• LLM provider decisions → "External Dependencies & Secrets Needed" section
• Coding conventions → "Coding Standards" section
• Behaviour rules for Claude → "How Claude Must Behave" section
• Infrastructure / deploy decisions → add a new "Key Decisions" section if needed

Keep CLAUDE.md under 200 lines. If it gets longer, tell the owner.

Step 4 — Mark applied

Update working_docs/session-log.md → change Status: captured to Status: applied for each entry written.

Step 5 — Summary

Tell the owner: "Applied {N} learnings to CLAUDE.md. Future sessions will follow these rules automatically."

---

When to use this skill proactively

Offer to run /learn log at the end of any session where:

• A decision was made about the tech stack or architecture
• A mistake was corrected (e.g. wrong config, wrong pattern)
• A reusable pattern was discovered
• The owner said "don't do X" or "always do Y"