Give Your AI Projects A Performance Review Like It's An Intern So It Stops Making The Same Mistakes
Hi Al friend!
You have probably noticed that Claude (or any other "intern" you're directing) gives you a solid first draft, and then you spend 15 minutes fixing the same three things you fix every single time.
Maybe you reformat the bullets into a numbered list. Maybe you cut a phrase you can't stand. And more often than you'd like, you re-paste your ministry's voice guide because the draft drifted formal again. It happens across many projects consistently, and it can feel like the tool never learns the small things. Most of us set up a project once and walk away, so it stays frozen at the same skill level while we keep doing cleanup a well-trained setup could handle on its own. Your setup can learn every bit of it, as long as you feed the lessons back in.
Today I want to walk you through the full approach for giving your Claude project a performance review, so it starts catching those fixes before you ever see them.
Here's how it works.
Stop settling for the first draft.
The first output is a starting point, and the real quality final version is ready after a few passes later.
Say you're drafting this week's midweek email. Pass one comes back generic. So you tell it to sound like a friend writing to the congregation, you ask it to name the actual sermon series, you cut the churchy filler. By pass four or five or seven(!) it might finally read like you wrote it.
Each of those corrections rounds are actually good data samples about how you actually want things to sound.
The habit you should build: keep going until it's right, and pay attention to what you keep asking for along the way. Every correction you say out loud is something you can later bake into the setup for good.
Write down the tweak you make twice.
After a few weeks inside a project, your patterns will start to show. You keep asking it to make the small group questions open-ended. Or the tone on the giving appeal always needs softening. Or your framework slips out of the draft so often you could set a watch by it. Those repeats are the raw material for the whole review, so keep a running note somewhere easy to reach, your phone or a doc, and jot the correction the second time you make it.
Two corrections for the same thing is enough to call it a pattern worth fixing. By the time you have 8 or 10 of these written down, you are holding a ready-made punch list for upgrading your entire setup.
Understand what the review actually does.
Once you have real history built up inside a project, you can ask Claude to study its own track record and report back on where it keeps missing.
It will read through every conversation you have had in that project, finds the corrections you have made over and over, groups them into clear patterns, and then hands you specific edits for your instructions so you stop repeating yourself.
Think of it as a performance review for your AI intern regarding their fundamental training or basic SOP's you've given them to handle.
Because I has watched every draft you produced, the AI itself can be the tool that tells you how to train itself better. Talk about inception for good!
You should run this loop every few weeks, once you have enough history for the patterns to be real.
Push the fixes into the source.
The review gives you edits, and now they need a home. So now what?
You have two places to put them:
- The first is your project instructions, which shape everything Claude does inside that project.
- The second is the skill file itself, if you built one, which is the underlying set of rules the project runs on.
As a general guide, put voice and formatting preferences into the instructions, and put anything structural about how the output gets built into the skill. After you do this, this is where you claw back the most time, because the source instructions now drive every output going forward.
Fix it once, and every future chat / output inherits the upgrade.
Treat your setup like an intern who is still learning.
A new intern has no idea about your pet peeves on day one. They learn by watching how you react over time.
Your Claude project grows the same way once you feed the lessons back into it. Run the review every few weeks or wait until you have a few dozen outputs in the project, which is enough history for helpful patterns to surface. Keep doing this and your setup will definitely get to know your voice, your frameworks, and all your ins and outs the longer it works with you.
It's the same with offline interns. -- all you're doing is the same thing for how a decent assistant turns into one you trust with the first draft.
Now for the part that makes this actually work. The whole approach hinges on two prompts: one that runs the performance review, and one that upgrades your skill file directly. I have both written out for you, ready to copy and paste, right below.
Grab both copy-paste prompts here ⤵️ ⤵️
