GTMcraft Plays · The Loop-and-Score Play

You run Claude one task at a time and grade by gut; build a loop that grades itself.

Play visual

claude-loop-and-score-signal.html

The 90-second read

You hand Claude one task at a time and grade the result by gut. So quality is a coin flip you re-run until it lands. The fix is a loop. You generate many candidates, score each against a metric you wrote, keep the winner, and log why the rest lost. Volume plus a scoreboard beats one hand-crafted try, every week.

• One handcrafted try is a slot pull; a scored batch of many is a workflow.
• Karpathy ran 700 experiments overnight and kept 20 real wins by scoring each.
• Eric Siu took the same loop to marketing, aiming past 36,500 tests a year.
• The lever is not the model; it is the metric you score against.
• A written scorer runs on a teammate's desk, so the loop stops needing you.

Read time: about 5 minutes. First step runnable Monday morning.

The #1 move this week

Turn one weekly output into a generate-5-and-score batch this Monday.

You already redo one asset every week and grade it by feel. Grading by feel means you are the scorer and you are not written down. So write the one number that defines a win, then have Claude produce 5 versions and rank them. Ship the top one, log the four that lost.

Start here: pick one output you redid 3 or more times last week. Write the single metric that says a version is good. You need that one output and one number, nothing else.

Ignore this week: building an overnight autonomous loop. Prove the batch by hand first, then automate the parts that hold.