Learn

Extract non-obvious session learnings into scoped AGENTS.md files to preserve knowledge across sessions.

When to Use

Activate after completing a non-trivial task to capture insights that would otherwise be lost.

Instructions

What to Capture (Non-Obvious Only)

• Hidden relationships between files or scripts not obvious from code.
• Execution paths that differ from what the code appears to do.
• Non-obvious config, env vars, or flags (see root AGENTS.md and agent_docs/building_the_project.md).
• Debugging breakthroughs where error messages were misleading.
• Files that must change together (e.g., AGENTS.md + agent_docs/AVAILABLE_SKILLS.md when adding skills).
• Build/test commands not documented in README.
• Architectural constraints discovered at runtime.

What NOT to Capture

• Obvious documentation or standard behavior.
• Duplicates of existing entries.
• Verbose explanations or session-specific notes.

Scoping Rules

Place learnings in the most specific AGENTS.md file:

• Project-wide: Root AGENTS.md.
• Script-specific: scripts/AGENTS.md.
• Skill-specific: .agents/skills/<name>/AGENTS.md.

Dual-Write Requirement

Every new non-obvious insight must be recorded in two places:

1. Verbose Log: Add a full LESSON-NNN entry to agent_docs/LESSONS.md with Issue/Root Cause/Solution.
2. Distilled Note: Add a 1–3 line note to the nearest AGENTS.md (this is what learn automates).

If agent_docs/LESSONS.md does not exist yet, create it before relying on the dual-write workflow.

Format

• 1–3 lines per insight in AGENTS.md.
• Fits within MAX_LINES_AGENTS_MD=150 constraint.
• Bulleted list under a "Learnings" or "Context" section.

Reference Files

• agent_docs/LESSONS.md - Project-wide lessons log.
• AGENTS.md - Root agent guidance and constraints.
• agent_docs/README.md - Index of workflow-impacting reference files.