I want you to help me set up my first graded loop: a way of having you check your own work against a bar I set, and redo it until it passes. We'll do it on one real task of mine. Follow the four parts below in order, and don't skip ahead to the task until I've confirmed the criteria.

How I want you to work: one question at a time, and wait for my answer. Reflect back what you heard in a line before moving on. If an answer of mine is too vague to grade against, say so and ask for the specific version. No flattery, no filler.

## Part 1: pick the task

Ask me these, one at a time.

1. "What's one thing you already ask AI to do more than once a week?" If I'm stuck, offer a couple of examples: pulling action items from a meeting, drafting a client update, summarising a report. If I name several, ask me to pick one.
2. "The last time you got that back, what was wrong with it?" Push past "it was generic". What specifically had to be fixed, or checked, or was simply made up?
3. "If you handed the output to someone you respect, what would they check first?" This is where the real criteria live.

## Part 2: write the criteria with me

From my answers, draft four to six criteria for what a good result looks like on this task. Each one has to pass this test: could you score the output against it without guessing what I meant? "Make it sound professional" fails that test. "Every claim links to a line in the source" passes, because it either does or it doesn't.

Write them as short numbered lines, each with a name and one sentence. Include at least one criterion about accuracy (nothing invented, nothing that can't be traced to the source) and at least one about completeness (nothing that should be there is missing).

Then show me the list and ask what's wrong with it. Change it until I confirm. Don't start Part 3 until I have.

## Part 3: run the loop

Ask me for the input: the transcript, the notes, the document, whatever the task works from. Then run three passes, and show me every step rather than doing it silently.

Pass one. Do the task the way you normally would. Label it "First draft".

Pass two. Grade the first draft against the criteria, and fix nothing. Lay it out as a table: each item or section of the draft in a row, and for each criterion a verdict of PASS, FAIL or NEEDS INFO. A criterion that applies to the whole draft rather than to one item, like completeness, gets its own row at the bottom. For every FAIL, name the single criterion it breaks and say why in one line. If something in the draft can't be traced to the input, that is a FAIL, and you quote the closest thing the input actually says.

Pass three. Fix everything that failed and give me a corrected version. Where you can't fix something because the input doesn't contain the answer, say so plainly. "No date agreed", "no owner named" and "not in the source" are correct answers. Never invent a value to make something pass.

Then grade the corrected version again, same table. If anything still fails, fix it and grade it again. Do this fix-and-grade cycle three times at most, counting the first correction as round one. If it still doesn't pass after the third round, stop and tell me exactly what failed and why.

## Part 4: what to keep

When it passes, or when you've stopped, give me two things.

The criteria, in their final form, as a block I can copy and paste at the top of my next prompt for this task.

One paragraph on what the grader caught that the first draft got wrong. This is the paragraph I read most carefully the first few times, before I let the loop run without me.

Then ask the first question from Part 1.