learning-capture

Learning Capture

Log corrections, errors, and discoveries to orbital/LESSONS.md so future sessions avoid repeating mistakes.

When to Use

Activate this skill whenever one of three triggers fires:

1. User correction -- the user says "no, do X instead" or "that's wrong." Log what you did, what the user wanted, and the corrected approach.

2. Tool failure recovery -- a tool call fails and you find a working alternative. Log the failing call, the error, and the fix.

3. Non-obvious discovery -- you learn something surprising about the project (e.g., a config file that overrides defaults, an undocumented API quirk). Log the finding and where you found it.

Approach

Read before act

Before starting any task type you have attempted in a previous session, read orbital/LESSONS.md with the read tool and scan for entries whose category or keywords match the current task. Apply any relevant lessons before your first action.

Entry format

Append one block per lesson. Use this exact structure:

### YYYY-MM-DD — <category>
**What happened:** <1-2 sentences>
**Do instead:** <1-2 sentences>
**Keywords:** <comma-separated tags for future search>

Categories (pick one): correction, error-recovery, discovery.

Writing the entry

• Use the edit tool to append to orbital/LESSONS.md. If the file does not exist, create it with the write tool.
• Keep each entry under 100 words. Precision beats length.
• Include enough context that a future agent session can act on the lesson without re-reading chat history.

Anti-patterns

• Do NOT log trivial observations. "The project uses Python 3.12" is not a lesson. Only log things that caused or would cause a mistake.
• Do NOT log every tool failure. Typos in a filename are not worth capturing. Only log failures where the recovery path was non-obvious.
• Do NOT duplicate entries. Read the file first; if a matching lesson already exists, skip or update it rather than appending a duplicate.
• Do NOT include full stack traces. Summarize the error in one line.