compound_skill

Overview

This skill captures and compounds knowledge from any session by extracting problems solved, decisions made, and gotchas discovered, then writing them to the task file for future agent retrieval. It runs after any phase or as a standalone step to improve future agents’ effectiveness by building a growing, searchable knowledge base. Use it to ensure lessons from debugging, design, and shipping are preserved and surfaced automatically.

How this skill works

The skill reads the specified task file, parses sections like research, plan, implementation, and ship, and scans conversation history to identify candidate learnings. It checks existing task notes and AGENTS.md for similar items to avoid duplication, structures findings into Problems/Decisions/Gotchas, appends a <future-agent-notes> section to the task file, and optionally promotes critical items into AGENTS.md as always-loaded context. It prompts before linking related tasks or promoting content so you control what becomes shared core knowledge.

When to use it

• After a ship or deployment to capture post-release findings
• After debugging sessions to record root causes and fixes
• When architectural or implementation decisions are made
• Whenever you encounter surprising or undocumented behavior
• Any time knowledge would save future rework or ramp-up time

Best practices

• Be specific: record concrete symptoms, root causes, and prevention steps
• Include alternatives and rationale for each decision so future agents can reason
• Search existing notes first to avoid duplicating prior learnings
• Only document meaningful learnings; skip trivial sessions with no surprises
• Use clear, self-contained items that make sense without full task context

Example use cases

• Capture a recurring database performance issue with symptoms, root cause, and index fix
• Record why a specific library or pattern was chosen and what alternatives were rejected
• Document a surprising platform quirk that broke tests and how to detect it early
• Append lessons after a feature ship so future agents start with known constraints
• Promote critical safety or deployment checks into AGENTS.md to always load for new runs