Hey, it's Cindy 🌱 You commented review, so here is the four-question after-action review adapted for Claude and ChatGPT.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
the outcome
What you get
A concise end-of-session review, an explicit keep-or-discard decision for every lesson, and a proposed patch that a human approves before anything persistent changes.
Use the guide in order. Gather the real inputs, run the prompts against those inputs, inspect the evidence, then make one reviewable change at a time.
start here
Set it up step by step
Run this before closing a session that produced a useful workflow, correction, or repeated mistake.
Answer the four AAR questions from the transcript: intended outcome, actual outcome, causes of the gap, and what changes next time.
Classify each lesson by destination: always-on instruction, reusable skill, example, source record, regression test, or temporary detail to discard.
Compare proposed changes with existing memory and rules so duplicates and contradictions are removed.
Review the diff. Save only durable, verified lessons, and keep private or one-off details out.
copy these
The prompts
End-of-session AAR
📋 copy this
Run an after-action review on this session. First answer four questions with evidence from the conversation and completed work: 1 What was supposed to happen? 2 What actually happened? 3 Why was there a gap, including assumptions, missing context, tool failures, or weak verification? 4 What should change next time? Then extract only durable lessons that would prevent repeated work or repeated mistakes. Classify each as: always-on instruction, reusable skill procedure, strong example, source record, regression test, or temporary detail to discard. Check the existing instruction and memory files before proposing changes. Remove duplicates, identify contradictions, and supersede anything this session proved wrong. Show a small exact diff and the destination for each change. Do not write anything until I approve it. Do not store secrets, private data, speculation, emotional venting, or one-off task state.
check the work
Quality gate before you trust it
1 Every saved lesson is reusable beyond this one task.
2 The proposed destination matches the lesson: rules stay short, procedures go in skills, failures become tests.
3 Existing instructions were checked for duplicates and contradictions.
4 No private, speculative, or temporary material enters persistent memory.
5 The final patch is small enough to review line by line.
6 Save the source links, as-of date, and final evidence beside the output so another person can reproduce the check.
Guardrail. The paper measures an average performance drop in simulated multi-turn tasks; it does not say models are intentionally designed to get worse. An AAR helps preserve lessons, but it does not guarantee future accuracy. Keep human approval on every persistent change.
keep it useful
Save the evidence with the workflow
Save the prompt with the source links and the date you verified them. When the tool, policy, price, interface, or documentation changes, update the guide before you reuse the workflow. Keep the inputs and outputs beside the project so the next run can be checked instead of reconstructed from memory.
the waitlist is now open
AI System Club is launching soon
Build AI systems with me through recorded lessons you can follow at your own pace. Join the waitlist for first access and the founding price.
btw this list gets first dibs the moment doors open :)