Most disappointing AI output isn’t a model problem. It’s a prompt problem, and the same five mistakes show up again and again when we review prompts submitted to the library.
None of them are dramatic. They are small habits that quietly cost quality. The good news: each fix takes less than a minute, and most of them are already built into the template approach we recommend.
1. Asking for everything at once
“Write a blog post, suggest five titles, create social posts and an email” is four tasks pretending to be one. The model spreads its attention and every part ends up average. Split it into separate prompts, or at least numbered steps with a clear output for each.
2. Hiding the format at the very end
If the output must be a table or JSON, say so near the top and repeat it at the end. A format instruction buried after three paragraphs of context is easy to miss, for people and models alike.
Return a Markdown table with exactly these columns:
| Feature | Benefit | Proof point |
Do not add any text before or after the table.3. Using vague quality words
“Make it better”, “more engaging”, “more professional”. These words mean something different to everyone. Replace them with something you can check:
- “More engaging” → start with a question the reader is likely asking.
- “Shorter” → under 80 words.
- “More professional” → no slang, no exclamation marks, no emojis.
4. Never showing what good looks like
One short example of a good output is worth a paragraph of instructions. It sets length, tone and structure in one go. Label it clearly so the model doesn’t copy it word for word:
Example of the style we want (do not reuse the content): "Cut onboarding from two weeks to three days. Here's the checklist our support team uses for every new account."
5. Never re-testing after a model update
A prompt that worked last spring might behave differently today. Models are updated, defaults shift, and a constraint that used to hold can start slipping. Keep a note of which model you tested with and when, and re-run your most-used prompts after any major update.
The prompt didn’t break. The ground under it moved.
Overheard in a team retro
The one-minute review
Before you save a prompt, read it once and ask: one task? Format stated early? Checkable constraints? An example? A test date? Five yeses and you’re ahead of most prompts we see.