Get Claude to mark its own homework

Why It Works

Think about what a good member of staff does before they knock on your door. They check their own work against what you asked for, and if it is not right they go back to their desk and try again, so you only ever see the version that passed. AI does not do that by default. It does the task once, gives it a quick read, and hands it back with the same spelling, the same tone and the same bit left out as last time, and you are the one who catches it.

The difference is a written test. "Is this good" is an opinion and Claude will happily agree with itself. "No American spelling" is a pass or a fail, and it can run that test on its own work as many times as it takes. Write down the things you always end up fixing, tell it to check against them before it shows you anything, and the loop takes over the part of checking you were doing by hand. You still read the result, you just stop reading it for the same three mistakes.

Step One, Write The Rubric With It

Here from the video? This is the line. You do not write the rubric yourself, because you already know what you keep fixing and Claude can get it out of you faster than you can list it. The prompt asks which setup you are in, interviews you one question at a time about the last few things you had to change, writes a list of three to six checks that each come out as a pass or a fail, and then runs the list on the last thing it wrote for you so you watch it catch something before you have to.

For a sense of what comes out, this is mine. Every word on this site is checked against it before I read a draft, and the video you came from shows it running.

1. 1British English, never American spelling
2. 2No emojis, anywhere
3. 3No colon in the middle of a sentence
4. 4No em dashes
5. 5No clipped sentence fragments, join the thought with and, because or so
6. 6No boxes drawn round text

Step Two, Use It On Any Task

Once the rubric exists in the conversation, one line at the end of any request makes that task check itself. Ask for the email, the plan or the document as you normally would, and add this underneath. It comes back already checked, with a note of what it had to fix, which is also how you find out what belongs on the list that is not there yet.

Step Three, Make It Apply To Everything

A line you have to remember to add is a rule you will forget on the day it mattered. In Claude Code the rubric belongs in the CLAUDE.md file in your folder, which Claude reads before every message, so the check runs on everything without being asked. In the Claude app the same words go in your project instructions and do the same job. This prompt puts it there and shows you exactly what it added, because a file it reads every time is not somewhere it should be writing unseen.

Step Four, A Skill For The Jobs You Repeat

Some jobs have their own checks on top of the general ones. A client email has to name the next step and a date, a proposal has to say the price once and never twice, a newsletter has to end on one door. A skill is how Claude Code keeps a job's own rubric and runs it every time that job comes up, and the app does the same with a saved project. This prompt asks which job, interviews you for the checks that only apply to it, and runs it once on a real one.

One Honest Bit

This only works for as long as there are real tests to go against. A rubric catches what you can name, spelling, a word you never use, a section that keeps getting left out, and it catches those every time. It cannot catch the thing you can only feel is off, and if you put "make it good" on the list you have handed Claude an opinion to agree with itself about. The taste is still yours, and the last read is still yours. What changes is that the read stops being about the same three mistakes.

Claude checking its own work is also weaker than a second pair of eyes checking it, because the same model that made the mistake is the one looking for it. For writing and admin that is a fair trade, the checks are simple and the cost of a miss is an edit. For anything where a miss is expensive, software that has to run, numbers that have to add up, the stronger version has a separate checker that actually runs the thing, and that is the verify loop.

Where This Goes Next

If the checks it keeps failing are about how it sounds rather than what it got wrong, the folder that writes like you is the fix underneath. If you want to understand what a skill is before you make one, a skill is a correction you kept is the plain version. And if your CLAUDE.md has grown to the point where the rubric is buried in it, the doctor finds what can go.