Prompt Iteration: How to Refine Until You Have Something Usable

A bad first output is not a reason to abandon the session. It is information. The output tells you what the model understood your brief to mean, which often differs from what you intended. Iteration is the process of closing that gap.

Most writers who are frustrated by AI writing are iterating incorrectly: they get a bad output, change the prompt randomly, and get a different bad output. Effective iteration requires diagnosing what went wrong before adjusting what to change.

Diagnosing the output

Four things can go wrong in an AI draft, and each requires a different fix.

Wrong structure means the piece is organized around the wrong argument or the wrong sequence. The model covered a topic rather than arguing a position, or it sequenced points in an order that does not serve the argument. This is a brief problem. The argument was not specified clearly enough in the brief, and the model filled the gap with topic coverage.

Wrong register means the piece is pitched at the wrong level. Too basic or too advanced for the target reader, too formal or too casual, too hedging or too declarative. This is usually a reader description problem. The brief specified a category of reader rather than a specific type of person in a specific situation.

Wrong specificity means the claims are general where they should be concrete. Statistics are missing or invented. Examples are hypothetical rather than real. This is a constraints problem: the brief did not specify that the piece should include specific claims or where those claims come from.

Wrong voice means the piece sounds like it was generated by a model. The hedging phrases, the symmetrical structure, the dead middle paragraph. This is a style reference problem. The system prompt either does not have a style reference or the reference is not specific enough.

Adjusting in the right order

Fix structure first. A structurally broken draft cannot be improved by voice editing. Identify what the argument should be, restate it in the brief, and regenerate. Do not iterate on a draft that has the wrong structure. Start over with the brief fixed.

Fix register second. If the structure is right but the pitch is off, add specificity to the reader description and regenerate. "A freelance content writer managing three to five clients" needs to become "a freelance content writer who is two years into the practice, managing four clients, producing eight to ten pieces per month, and looking to take on two more clients without increasing working hours."

Fix specificity third. If the structure and register are right but the claims are generic, add the specific claims you want to the brief. Do not ask the model to be more specific in the abstract. Give it the specific information and tell it to use it.

Fix voice last. Voice is an editing pass, not an iteration problem. If the structure, register, and specificity are right but the piece sounds AI-generated, run the editing pass rather than iterating on the prompt. Iteration does not reliably fix voice. The editing pass does.

When to give up on a session

If you have iterated twice on the structure and it is still wrong, the brief is not fixable within the session. Stop, go back to the brief, and rewrite it from scratch. Sometimes the act of writing a brief reveals that you do not yet have a clear enough argument for the piece to be written.

If you have iterated twice on the register and the pitch is still off, the reader description needs a complete rethink. Write three sentences about a specific real person in the target audience. What did they search to find this piece? What do they already know? What do they need to be able to do after reading it? Use those sentences as the reader description.

If the voice never improves through iteration: it will not improve through iteration. Voice is corrected in editing, not in prompting. Accept that the prompting stage has a voice ceiling and plan the editing pass accordingly.

What iteration is not

Iteration is not adding more words to a failing prompt. A longer prompt applied to a broken brief produces a longer version of the same broken output.

Iteration is not asking the model to improve the draft in general terms. "Make this better" or "make this more engaging" produces a different version of generic output. The improvement request has to be specific: "the opening paragraph describes what the piece covers rather than opening with the problem. Rewrite it to open with the problem."

Specific feedback on a specific problem produces specific improvement. General feedback produces general improvement, which is another way of saying it produces the statistical center of what improvement looks like, which is not usually what you wanted.