Learnt — Session Teaching Loop

Review the current conversation and extract non-obvious lessons worth keeping. The goal is to make the next session smarter, not to record everything that happened.

What's worth capturing

Look for:

• Corrections — Claude did X, user said "no, do Y instead"
• Validated approaches — user confirmed or accepted a non-obvious choice without pushback (quiet wins count)
• Revealed preferences — how the user likes to work, what they care about, what annoys them
• Project facts — decisions, constraints, context that isn't in the code or git history

Ignore:

• Things derivable from reading the current code
• Standard best practices Claude already knows
• One-off task details with no future relevance
• Anything already in CLAUDE.md, MEMORY.md, or the global skill files

Decision tree — where does each lesson belong?