The most common prompt engineering mistakes are being vague, skipping examples, mixing notes with instructions, asking for the whole book in one go, and judging a prompt by whether the answer sounds good instead of whether it meets a stated requirement. Anthropic’s prompting best practices start from a useful test: show your prompt to a colleague who lacks your context, and see whether they could follow it. If they would be confused, the model will be too. OpenAI’s prompt engineering guide describes prompting as a mix of art and science because outputs are non-deterministic, which is why a single lucky answer is not proof that a brief is solid. Google’s prompt design strategies treat prompt design as iterative: experiment, rephrase, and refine based on the results you actually get.
For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up, so you can catch several of these mistakes at the system level—separate stages, a standing brief, and a review step—rather than only in a one-off chat. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article, not paid Builder prompts and not recorded model outputs.
Mistake 1: asking for a topic instead of a job
A vague prompt:
Write a chapter about client boundaries.
A better prompt names the reader and the job:
Write a 300-word opening plus a five-heading outline for new freelance
designers who say yes to extra revisions because they are afraid of
seeming difficult. The chapter's job is to help them state a revision
limit in the kickoff email, not to draft a legal contract.
If a colleague could not tell what “done” looks like, the brief is still a topic. Anthropic’s “brilliant but new employee” framing is the right standard here. For more starter rewrites, see prompt engineering examples for beginners.
Mistake 2: leaving the notes in your head
The model cannot use the example, constraint or conflict you forgot to paste. Google’s guidance is explicit: include the information the model needs, instead of assuming it already has it. A before-and-after makes the miss obvious:
Write the section on the three morning actions. You know the ones I mean.
Draft the section "Choose three actions" using only these notes:
- Review today's calendar
- Write the first task in one sentence
- Prepare the workspace for the first call
The reader has 45 minutes, three days a week. Do not add other habits.
If a detail is missing, write "Need from me."
If you have more than a few lines of source material, put it in a labeled block. Our guide to delimiters covers that.
Mistake 3: describing the vibe and skipping the format
“Make it engaging,” “make it professional,” and “don’t be generic” are almost impossible to check. OpenAI, Anthropic and Google all ask you to specify the output format. Replace the vibe with a shape:
Return 250 to 350 words of teaching prose, then three numbered steps
with a time estimate for each. End with one workbook question and two
blank lines for the reader's answer.
Keep a short tone note if you need one—“calm and practical, no hype”—but let the format carry the test.
Mistake 4: skipping examples when the shape keeps drifting
If you have already written a clear brief and the headings, voice or workbook pages still wander, you probably need a sample. Google recommends examples for regulating format, phrasing and scope. Anthropic calls examples one of the most reliable steering tools. OpenAI describes few-shot learning as showing a handful of input/output pairs. The mistake is not “using examples.” The mistake is using none after the same miss has already happened, or dumping an entire old chapter into the prompt and saying “like this” without marking what “this” is. See zero-shot versus few-shot prompting for the decision, and the few-shot chapter guide for the book-specific method.
Mistake 5: mixing instructions, notes and old drafts into one unlabeled pile
When everything is pasted together, the model may follow the wrong part. Use headings or tags, and put the request last, especially when the notes are long. Anthropic reports that placing the query at the end can improve quality by up to 30 percent in tests on complex, multi-document inputs. Google’s Gemini 3 guidance also says to put large context first and the question at the end. The mistake is not pasting context. The mistake is hiding the job inside it.
Mistake 6: asking for the whole book in one request
A manuscript is several jobs: plan, draft, design the exercises, and edit. Google advises breaking complex work into components and chaining prompts when the steps should be sequential. A one-shot “write my book” request buries every decision in a single pass, so you cannot see where a problem started. Split the work. Our guide to prompt chaining and our guide to prompts for long-form content describe that structure. In Interactive eBook Builder, Spec Builder, Drafter and Polisher exist for the same reason.
Mistake 7: using a costume instead of a brief
“Act as a world-class expert” does not supply notes, does not set a format, and does not make unsupported claims true. A short, specific role can help with audience and tone. A costume tries to replace the rest of the prompt. If you use a role, keep it modest and still include the job, the notes and the limits. Our article on whether you should tell AI to act as an expert and our comparison of role prompting and system prompts cover the distinction.
Mistake 8: judging the prompt by whether the answer sounds good
This is the mistake that hides all the others. A fluent section can still invent a rate, skip the workbook page, or write for the wrong reader. OpenAI emphasizes that you should not treat one good-looking response as proof, because generation is non-deterministic. Anthropic starts with success criteria and a way to test them. Build a short checklist from the brief—reader job, notes used, format met, bans respected—and score the result against that, not against polish. Then change one variable and try again. Our guide to improving a prompt when AI is generic is the longer version of that loop.
A related error is changing five parts of the prompt at once after a miss. You will not know what helped. Save the previous version.
Mistake 9: letting a standing brief rot
A template or system prompt is a mistake in waiting if you never reread it. It can accumulate contradictory rules, leftover notes from chapter 1, or a ban list nobody can apply. Keep standing instructions short enough to review. Move chapter-specific material into the request. Change the standing brief only when the same miss appears more than once, or when a claim would be harmful even once.
Mistake 10: treating a first success as a finished method
Try the same prompt on a second chapter or a second set of notes before you reuse it across the book. Google’s iteration section exists because the first phrasing is often not the one you want to keep. If you are building a product around the brief, test it the way you would test any other working document: on more than one input, with the same requirements.
A quick map of the misses
| Mistake | What you asked | What to ask instead |
|---|---|---|
| Topic, not a job | “Write about boundaries” | Name the reader and the job the page must do |
| Notes left in your head | “You know the ones I mean” | Paste the notes in a labeled block |
| Vibe, no format | “Make it engaging” | State length, structure and what the page includes |
| No examples after drift | A clear brief that still wanders | Add one or two approved samples of the shape |
| Unlabeled pile | Rules, notes and old drafts mixed together | Delimit the parts and put the request last |
| The whole book at once | “Write my 12-chapter eBook” | Plan, draft and edit in separate, inspectable steps |
| Costume instead of a brief | “Act as a world-class expert” | A modest role plus notes, format and limits |
| Judging by polish | “This sounds good” | Score the result against the stated requirements |
| A rotting standing brief | Contradictory rules you never reread | Keep the policy short and review it |
| One lucky success | Reusing a prompt after a single good run | Test it on a second set of notes first |
What a healthier loop looks like
- Write a brief a colleague could follow.
- Include the notes and a format you can inspect.
- Run it once, then score the result against the brief.
- Revise one part.
- Run it on a second set of notes.
- Only then save it as a template.
Spec Builder, Drafter and Polisher give you a place to run that loop at each stage. SEO BOT can later publish supporting articles to the Auto-Blogger blog, each linking back to your eBook, site and brand, but those articles need the same checks or you will scale a fluent miss. Keep or adapt the supplied setup, and review before anything goes live.
FAQ
I already use a popular framework. Can I still make these mistakes?
Yes. A framework only helps if each part contains a decision. An empty “Objective” or “Context” heading is still a vague prompt. See prompt engineering frameworks explained.
The model ignored a rule I wrote in all caps at the bottom. What now?
Move the rule into the standing instructions, say it in a positive format if it is really about shape, and keep the request last. Emphasis at the end of a messy brief is weaker than a short, early, testable instruction. Then inspect the next result against that one rule.
When is a short prompt not a mistake?
When the task is small and the missing decisions are genuinely obvious, such as “Turn this approved outline into a title-case H2 list, no new items.” Short is not the error. Missing a decision the model will have to guess is the error.
How many times should I test a prompt before I reuse it?
There is no official number. A practical minimum is two realistic inputs, because one success can be luck. If the book has several page types, test one of each type.
Before you hit send
- Could a colleague follow this brief without the context in your head?
- Are the notes in the prompt, labeled, and separated from the request?
- Have you described a result you can inspect, not only a mood?
- Are you asking for one job, or for a whole book at once?
- Will you score the output against the brief, then change only one variable?
- Have you tried the same prompt on a second set of notes before standardizing it?