← ClaudeAtlas

auto-improvelisted

After completing any task, extract reusable rules from what worked and what failed. Produces output in the exact format used by agent.instructions.md (## Mandatory: Rules to Follow / ## Mandatory: Rules to Avoid). Trigger after a task finishes when the user asks for a retrospective, lessons learned, or rule extraction — or implicitly with phrases like "what did we learn", "what went wrong", "what should we avoid next time". Do NOT trigger mid-task or before execution completes.
Ronnasayd/aipp · ★ 3 · AI & Automation · score 62
Install: claude install-skill Ronnasayd/aipp
# Task Retrospective A skill for reflecting on completed tasks and converting experience into lasting rules. The goal is simple: after finishing work, extract what was learned — both what went well and what failed — and encode that knowledge as concrete, actionable rules in the same format used by `agent.instructions.md`. This prevents repeating the same mistakes and reinforces effective patterns. --- ## When to use this skill - Immediately after completing a task, feature, plan, or agent workflow - When the user asks for a reflection, retrospective, or lessons learned - When an agent wants to close a loop and capture institutional knowledge - Proactively, at the end of any multi-step execution that involved meaningful decisions --- ## Process ### Step 1: Collect context Gather the raw material for reflection. Use whatever is available: - Current conversation history (decisions, corrections, feedback given by the user) - Task/plan files if provided (e.g., `docs/agents/plans/*.md`, `tasks/todo.md`) - Git diff or commit history if relevant - Any error messages, failed attempts, or corrections made mid-task If a file path was passed as argument, read it. If nothing was provided, analyze the conversation directly. ### Step 2: Identify decision events Go through the execution timeline and identify discrete moments where a decision was made. For each one, note: - **What was decided** — the action taken or approach chosen - **Outcome** — did it work? Was it corrected?