THE LOOP METHOD
AN EXAMPLE PROMPT
This week's AI news, turned into something you can use today. The skill in 2026 isn't a cleverer single prompt, it's telling the model how to check its own work, when to retry, and how it knows it's actually done.
How the loop works
Most people only ask the model to do the work once, then stop. This template forces it to also judge its own output against criteria you set, retry when it falls short, and know exactly when it's actually finished. Works for anything with more than one step and a definition of good: emails, spreadsheets, code, research, event copy, ad copy, anything.
CHECK›
FIX›
DONE
Fill in three blanks
GOAL, your CHECK criteria, and your DONE WHEN condition. That's the whole setup.
Paste it whole
Into any AI tool you're already using. No special setup required.
Let it run
Don't interrupt after step 1. The point is it checks and fixes its own work before you ever see it.
Make the check hard
"Make it good" is not a check. Specific, testable criteria are.
Runs anywhere you already work
Nothing in this prompt depends on tool calls, file access, or a specific product. It's plain language about how to work, so it runs the same way everywhere. Change the harness, keep the loop.
Terminal-based agentic coding. The loop is exactly how it already works task to task.
Desktop agent mode for files and tasks. Same loop, applied to documents and everyday work.
Any chat interface. Paste the prompt, let it interview and refine before it hands you anything.
No product at all, just a system prompt. The loop is instructions, not a feature.
The template
Fill in the three blanks in brackets, then paste the whole thing in.
GOAL: [what you want when this is finished]Work in a loop, don't stop after the first pass:1. ATTEMPT. Do the next piece of the goal.2. CHECK. Before moving on, test what you just did against these criteria: [specific, checkable criteria, not "make it good"]3. IF IT FAILS THE CHECK. Diagnose what's wrong, fix it, and check again. Do not move to the next step until this one passes.4. IF IT PASSES. Move to the next piece of the goal, then repeat step 2.5. DONE WHEN: [a specific, testable definition of finished]. Don't stop early just because you're out of obvious next steps. If you believe you're not done and don't know why, say so and ask what's blocking it, don't guess and stop.Repeat steps 1 through 4 until the condition in step 5 is true. Then tell me it's done and why.
GOAL: Three follow-up emails for people who toured Workuity but haven't booked a space yet. Each one should get a reply, not just a read.Work in a loop, don't stop after the first pass:1. ATTEMPT. Draft email 1.2. CHECK. Read it back as the visitor would. Does it sound like a person, not a template? Is there exactly one clear ask? Is it under 150 words?3. IF IT FAILS THE CHECK. Rewrite it, then check again.4. IF IT PASSES. Move to email 2, then email 3, repeating the same check each time.5. DONE WHEN: all 3 are drafted and checked, none use a hedge phrase like "just checking in" or "wanted to follow up," and each takes a different angle (tour recap, social proof, urgency).Repeat steps 1 through 4 until the condition in step 5 is true. Then tell me it's done and why.
Why this works
Separates do from judge
Most people only ask the model to do the work. This forces it to also judge its own output before you ever see a draft.
The check has to be checkable
"Make it good" is not a check. "Under 150 words, one clear ask, no hedge phrases" is. Vague checks produce vague loops.
Done When kills the guessing game
Without it, the model decides on its own when to stop, usually too early. With it, you've defined finished before it starts.
A skill is just this, saved
Every Claude skill Dan uses in Cowork is this same pattern: a goal, a way to check the work, a definition of done, named so you don't retype it.