The Claude Rewrite Loop Pros Use
A simple way to get A++ outputs everytime.
The pros never accept Claude’s first output. They make the model grade its own work against a rubric, beat the grade, and do it more than once.
In this post, I will give you:
🎯 The rubric tactics pros use to get Claude to go above and beyond.
🔁 The four-round method that makes each pass grade harder than the last.
🎁 My copy-and-paste meta prompt loop you can run on anything.
Why the second draft beats the first
It is far easier to taste that a soup needs salt than to cook a perfect soup from scratch. Spotting a flaw is easier than avoiding one. So when you ask Claude to stop, taste its own work against a real standard, and fix what it finds, it catches things the first pass missed.
How much does this actually help? The research that first tested this loop had a model write, critique itself, then rewrite, with no human in the seat. Quality climbed about 20 percent on average across a range of tasks. That is a real jump for one extra message.
Build a rubric the model has to pass
At the end of this post, I list my meta prompt rewrite loop that does it all (including making the entire rubric). Even if you use my rewrite loop, it helps to understand why the rubric matters below.
A rubric is just a grading checklist, the kind a teacher uses to mark a test.
Tell Claude “rate this 1 to 10” with no rubric checklist, and you are handing a student their own test with no answer key. They give themselves a 9, point at the parts that were already fine, and move on. The number means nothing, because nothing got checked against a standard.
The trick to a good checklist are questions that can only be answered yes or no.
Vague questions let the model wriggle free. Sharp ones pin it down.
Here are some bad questions made better.
⛔ “Is it engaging?”
✅ “Does the first line make you want to read the second?”
⛔ “Is it clear?”
✅ “Could a stranger follow this with no follow-up question?”
⛔ “Is it persuasive?”
✅ “Does every claim have a reason sitting under it?”
A question you can answer yes or no is a question the model cannot argue its way around.
5 rules of a good rubric
Here are five rules that turn a soft checklist into one that actually holds.
#1 Make it show its work before the grade. Tell the model to write the reason first, then the numeric grade. A bare number is easy to fudge. Forcing the explanation first is like making a student show the math before the answer, so the soft spots have nowhere to hide.
#2 Put one hard question on the test. Add a criterion the work has not earned yet, so there is always somewhere left to climb.
#3 Make the questions that matter more worth more points. If a weak opening should sink the whole piece, say so. Tell the model to cap the score at a 6 whenever the first line is flat.
#4 Let it leave a box blank. Tell the model it can answer “can’t tell” when it does not have enough to judge, rather than inventing a confident number. A grader allowed to say “I don’t know” is one you can trust on the boxes it does check.
#5 Grade one question at a time. For anything that matters, score each item on its own instead of blending everything into a single number. One overall grade quietly averages the real problems away, the way a B+ can hide one failing section.
Make the judge meaner each round
Raising the bar every round is what separates a tidy edit from a real climb. Each pass should face a tougher judge than the last.
Raise the threshold.
A 10 from the last round counts as a 7 in this one.
Harden the persona.
A neutral grader becomes a tough editor becomes the reviewer nobody satisfies on a first try.
Before it re-scores, make it argue against its own draft.
Tell it to name the three best reasons this is still not a 10. A model talked into hunting for faults, grades itself far more honestly than one asked whether its work is good.
How many rounds should the loop go
When you run a rewrite loop the gains come fast in the beginning. By the completion of a fourth round, you are entering the 99th percentile.
Round 1 does most of the work.
Round 2 raises the floor.
Round 3 cuts what does not belong.
Round 4 is polish.
Always tell it to stop after the fourth round. Past four rounds you are just wasting tokens.
Here’s my meta prompt rewrite loop
This meta prompt is setup to do all of the following for you:
Researches to find what excellence is.
Builds a full rubric for grading.
Creates a four round loop (each tougher than last).
Delivers new output with notes on what changed.
Below is the copy and paste prompt that you run after you get your first draft back from Claude. Only paid members have access to the meta prompt below. We’re currently running a 75% off special, so now is a great time to upgrade.