The self-review hack that turns any Claude skill or workflow into something that gets better every time you run it. Five minutes of upkeep, compounding returns.
Most people's Claude prompts are frozen in time. You wrote them in March, you've run them 80 times since, and they are functionally the same prompt today as they were the day you wrote them. That is wild when you think about it. The whole point of working with AI is that it gets better when you give it better instructions, and instructions get better with feedback.
This guide gives you the exact 3-question review block I add to the bottom of every prompt I rely on, plus the simple weekly habit that keeps the prompt compounding instead of decaying.
When you first build a prompt, you tune it. You run it five times, notice it keeps missing the brand voice, you add a brand voice paragraph. You notice it gives generic openings, you add three example openings. You iterate hard for the first week.
Then it gets good enough and you stop touching it. From that point on, the prompt only gets worse relative to your actual needs. Your business evolves. Your taste sharpens. The model itself gets updated. But the prompt sits there frozen.
The fix is structural, not motivational. You bake the upgrade loop INTO the prompt so it happens automatically, not whenever you remember to look at your prompts (you won't).
Paste the block below at the bottom of any Claude prompt, skill, or system instruction. Once it's there, every single time Claude finishes the main task it will automatically produce a 3-line review of its own output before handing it to you.
Copy this prompt
At the end of every response, append a section called REVIEW that contains exactly three bullet points: 1. WHAT WORKED. The strongest part of what I just produced and why. Be specific (cite the line, the choice, or the input that made it work). 2. WHAT WAS WEAK. The part of the output that felt the worst. Where I had to guess, fall back on generic language, hedge, or skip a real answer. Be honest, not diplomatic. 3. WHAT TO ADD TO THIS PROMPT OR THE KNOWLEDGE BASE BEFORE NEXT RUN. The single highest-leverage change to the prompt, the example set, the system instructions, or the project knowledge files that would have made this run noticeably better. Phrase it as a concrete addition I can paste in. Always include the REVIEW section. Never skip it. Never make it sound like marketing copy. Treat it like an honest post-run debrief between you and me.
Why three, not five
I tested 5-question versions, 7-question versions, and structured rubrics. They all bloat the output and stop being useful by the third week. Three is what survives at scale. Keep it tight.
The block is half the system. The other half is what you do with the reviews.
Once a week, open the prompts you use the most. Scroll through the REVIEW sections from that week's runs. Most will be minor. Skip them. About one in five will have a genuinely useful suggestion: a missing instruction, a gap in your context file, a specific example you should add.
Approve the good ones and paste them in. Either into the prompt itself, into the project's CLAUDE.md or instructions file, or into a knowledge file Claude has access to. Save. Move on.
That is it. Five minutes a week per prompt. After a month, your prompt is genuinely smarter than the day you wrote it. After three months, it is unrecognizably better.
If you use Claude Cowork or Claude Projects, do not paste this block into every prompt manually. Build a reusable Skill called 'Self-Reviewing Output' whose only job is to append the REVIEW block to whatever main task you've asked for.
Then you can invoke it from any chat with one line: 'Run this with self-reviewing output.' Claude knows what that means and applies the block automatically.
The hidden compounder
The reason this works so well is that it removes the activation energy. You are not asking yourself to remember to review prompts. You are asking Claude to hand you the review on a plate, every single time. The hard part (figuring out what's wrong) gets done automatically. You only do the easy part (deciding whether to act on it).
The Only AI Masterclass You Need
If this guide helped, but you’re looking to go deeper, I got you!! My 30-Day Challenge takes you from saving AI tips you never use to actually building with AI, step-by-step.
I show you exactly how I automated two e-commerce brands, my social media, and most of my personal life, then hand you the agents, workflows & systems to do the same. I’m teaching you every single thing I know with one lesson and one build a day.
Join the AI Masterclass →© 2026 Mariah Brunner. All rights reserved.