Improve an answer through iteration
Iteration in POM AI is not a conversation. There is no chat thread, no memory between runs and no "try again but shorter". Each generation starts from scratch, reads your prompt and your configuration, and charges credits.
That changes how to iterate efficiently.
The two free levers
Estimates are free. You can price a configuration as often as you like, including to compare two before choosing.
Editing is free. Cutting a generated draft, fixing a heading, deleting a section — all free. Regenerating to avoid ten minutes of editing is usually the wrong trade.
Before iterating, ask whether editing gets you there. Frequently it does.
Change one thing at a time
Because each run is independent, changing three things means you learn nothing about which mattered.
A disciplined loop:
- Generate.
- Name the single biggest problem.
- Change one element of the prompt or one control.
- Generate again.
- Compare.
Two or three rounds is normal. If you are on the sixth, the problem is upstream — usually the business description or a target length the topic cannot support.
Iterate at the cheapest stage
The multi-stage tools have cheaper planning work and more expensive output work:
| Tool | Cheap stage | Expensive stage |
|---|---|---|
| Long-form post builder | The plan | Every section, optional image and SEO metadata |
| Site-wide copy generator | The page plan | Every page |
The long-form builder does not pause to expose an editable plan. Review the prompt, target length, profile and image maximum before accepting the first-step estimate. For a completed article with a weak structure, small changes are cheaper in the editor; substantial changes require a better prompt and a new run after the current generation is finished. See Prepare a long-form article plan.
The site-wide tool goes further: its plan is directly editable, so you can fix the structure by hand for free rather than regenerating it.
Iterate on something small
Test a prompt idea on the smallest thing that will show whether it works:
- one paragraph, not one article;
- one image at a lower resolution before processing a batch;
- one translated post, not the whole site;
- ten images for metadata, not the library.
Once the approach is right, scale it up.
Diagnosing what to change
| The result is | Change |
|---|---|
| Generic and interchangeable | The business description, not the prompt |
| Off-topic | Add must-have points and exclusions |
| Right topic, wrong reader | Add a sentence describing the reader's situation |
| Too long or padded | The word count control, not the prompt |
| Wrong tone | The copywriting style control, then the business description |
| Fragmented into shallow sections | Prefer fewer, deeper sections in the prompt |
| Repetitive across sections | The outline — two sections overlap |
| Missing something you assumed | Say it explicitly; nothing is implied |
The first row is the most common and the most often misdiagnosed. Prompting harder does not fix a thin business description; six unsatisfying attempts in a row is the signal to go and edit that instead.
Keep what works
The moment a prompt produces a good result, save it under Settings → POM AI → Prompt templates, assigned to the contexts where it applies. Otherwise the same discovery gets re-made, badly, by whoever generates next.
Templates append to the prompt field, so a saved skeleton can be applied and then adapted for the specific case. See Use saved prompt templates.
When to stop
Stop when you are choosing between preferences rather than fixing problems. At that point the remaining gap is editorial work — adding your own examples, opinions and specifics — and no further generation will produce it.
Stop and reconsider entirely when three attempts have each been wrong in a different way. That pattern means the request itself is unclear, not that the wording needs another pass.
Cost of iteration
Every generation charges. Iterating three times on an article is roughly three articles' worth of credits.
The practical protections: make plan requirements explicit before starting a long-form article, iterate on single paragraphs rather than whole pages, and use the least expensive profile that meets the task. See Understand credit estimates.