Make Claude Grade and Fix Its Own Work

The Forge Loop is one standing instruction you paste into Claude once. After that, every task runs the same way: Claude does the work, turns around and grades its own work like a harsh critic, fixes what it finds, and repeats until the work actually passes. You stop being the QA department.

Where to paste it: in the Claude app, open your Project, then Instructions, and paste Prompts 1 and 2 together. In Claude Code, paste them into your CLAUDE.md file. Same words work in both.

This is the engine. It forces the turn-around moment: the same model that wrote the draft has to attack it like a stranger wrote it, then fix only what it found.

After you finish any task, do not stop. Run the Forge Loop:

1. 1
   Switch roles. You are now a harsh critic reviewing this

work like a stranger wrote it. Grade it with my rubric,

score it out of 100, and list every specific issue.

1. 1
   Fix only the issues you listed. Do not rewrite the parts

that already work.

1. 1
   Grade the fixed version again.
2. 2
   Repeat until the score is 90+ with zero open issues,

or you have done 5 loops.

Show me one line per loop: "Loop N: score/100, what changed."

Then show only the final version.

A loop is only as honest as its rubric. Without one, Claude gives itself a polite 92 and stops. This one makes the grade mean something. Edit the five lines to match what you care about.

Grade out of 100 across these five, 20 points each:

• Correct: no factual errors, nothing made up, numbers check out
• Complete: nothing the task asked for is missing or half done
• Clear: a first-time reader gets it, no filler, no jargon
• Consistent: matches my voice, my formats, my past examples
• Tested: edge cases checked (empty input, weird data, long text)

Anything under 18/20 in a category is an issue you must list

and fix. Be harsher than feels fair. A polite 85 helps no one.

This is what makes it hands-off. Without it, Claude asks permission between loops and you are babysitting again. With it, you send one message and come back to finished work.

Run [THE TASK]. Then apply the Forge Loop from my instructions.

Do not ask me anything between loops. Do not show me drafts.

Stop only when: score is 90+, zero open issues, and the work

passes its own check twice in a row. If you hit 5 loops and

still fail, stop and show me the top 3 blocking issues instead.