Iteration: Prompting Is a Loop, Not a Shot
A working note on the most important habit the other sheets assume: reading what comes back and adjusting.
Most prompting advice, including the other sheets in this series, talks about the prompt as if you fire it once and judge the result. Real work doesn't happen that way. The skill that separates fluent users from people reciting tips is treating the exchange as a loop: read what came back, diagnose why, change the cause, not the symptom.
The common move
This treats a weak output like a slot machine: pull again and hope. Sometimes it improves, often it doesn't, and you learn nothing either way. The question worth asking: why did it come back the way it did?
Pulling the lever, versus reading the output
- "Try again" with no diagnosis. You re-roll without knowing what was wrong.
- Treating output as pass/fail. Good or bad, with nothing learned about why.
- Adjusting the symptom. "Make it shorter, warmer, smarter" without asking what produced the miss.
- Diagnosing the cause. Was the context thin? Did the framing demand confidence? Did the examples imply the wrong thing?
- Treating output as evidence. Each response tells you something about your prompt.
- Adjusting the cause. Fix what produced the miss, not just its surface shape.
A bad output is not a failure to discard; it's information about your prompt. The loop is read, diagnose, adjust, and the diagnosis is the part most people skip.
The more honest move: diagnose before you re-prompt
When an output misses, resist the reflex to re-roll. Ask what in your prompt produced this specific miss, then change that one thing.
A short diagnostic vocabulary, drawing on the other sheets: too generic? Your instruction was abstract; show an example instead (sheet 6). Confidently wrong? Your framing demanded an answer; make uncertainty safe. Missing what you meant? The context was thin; add what you left out. Wrong shape? You described instead of demonstrated.
Notice the difference from "make it better": you named the likely cause, changed it, and asked the model to surface what it inferred, so the next output also tells you whether your diagnosis was right.
Try this
When an output disappoints, don't re-roll. First say, in one sentence, why you think it came back that way.
Change the one thing your diagnosis points at, not five things at once. Changing everything teaches you nothing.
Read the next output as a test of your diagnosis, not just as a new attempt.
The principle underneath
A weak output isn't a dead end; it's evidence about your prompt. Diagnose why it came back as it did and change that, instead of pulling the lever again.