A prompt should be as long as the task needs and no longer: a line or two for a quick fact, roughly 100 to 400 words for a chapter draft, and a clearly labeled block when you paste in source material. The word ranges are my own rule of thumb, but the "no longer" part has published evidence behind it: in controlled tests, models reasoned worse as irrelevant text piled up, and they used information in the middle of long inputs less reliably than information at the start or end.
Below you'll find the numbers from two studies, a worked chapter prompt with a measured word count, where to put each part, how many examples to include, and a trim test you can run in about ten minutes.
How long should a prompt be for an ebook chapter?
For most chapter drafts, aim for 100 to 400 words, which is roughly 135 to 535 tokens. That is enough room for the reader, your notes, a voice sample, a few rules and the task, and short enough that every line still does something. OpenAI's token guide gives the conversion for English text: about 100 tokens per 75 words, or about 4 characters per token.
| Task | Prompt length (my rule of thumb) | What earns its place |
|---|---|---|
| Fact, definition or rewording | 10 to 40 words (about 13 to 53 tokens) | The thing to define and who it's for |
| Brainstorm | 20 to 60 words (about 27 to 80 tokens) | Topic, reader, how many ideas, "rough is fine" |
| Edit pass | 80 to 200 words (about 107 to 267 tokens) | Edit level, what to protect, how to list changes |
| Chapter draft | 100 to 400 words (about 133 to 533 tokens) | Reader, voice sample, notes, 3 to 5 rules, the task |
| Chapter from pasted research | 100 to 400 words of instructions, plus the source | Source first, each piece labeled, task last |
Here is a chapter prompt for a budgeting workbook. I counted it: 113 words, about 151 tokens by the 100-per-75 rule.
<context>Readers: freelancers with irregular income who dislike spreadsheets. Book: a 12-chapter budgeting workbook. Tone: plain and friendly, no jargon.</context>
<source>My notes: pay yourself a fixed monthly "salary" from a separate account; keep one month of expenses as a buffer; set aside a tax percentage the day each invoice is paid.</source>
<example>"Open your business account and look at last month. Do not judge it. Just write down what came in and what went out."</example>
<rules>1. Second person, short sentences. 2. No tax or legal advice beyond "check with a tax professional". 3. End with one worksheet task. 4. 700 to 900 words.</rules>
<task>Write Chapter 3, "Pay yourself a salary", using only the notes above.</task>
Look at what is missing. There is no "you are a world-class author" preamble, no "please be accurate" filler and no rule that appears twice. Each block answers something the model can't guess: who the reader is, which facts to use, what good sounds like, where the limits are and what to produce. A role line can earn a place when it's specific, which we cover in whether to tell AI to act as an expert.
Do longer prompts give better results?
Not automatically: extra context helps only while it's relevant, and past a point extra text makes answers worse. Two studies put numbers on it.
In Same Task, More Tokens (2024), Mosh Levy, Alon Jacoby and Yoav Goldberg took simple True/False reasoning questions that the models mostly answered correctly, then padded each one with irrelevant text up to 250, 500, 1,000, 2,000 and 3,000 tokens. The facts needed never changed. Averaged over the five models tested, accuracy fell from 0.92 to 0.68 at 3,000 tokens, which is about 2,250 words and far below those models' advertised limits. Asking for step-by-step reasoning didn't close the gap for most of them, and longer inputs made models more likely to ignore the instruction to answer True or False.
One caveat: the models were GPT-4, GPT-3.5, Gemini Pro, Mistral Medium and Mixtral 8x7B, so newer models may hold up better. Treat the result as a reason to test your own setup, not as a law.
The second study, Lost in the Middle (Liu et al., from Stanford, UC Berkeley and Samaya AI), gave models 20 passages and asked a question that exactly one passage could answer, then moved that passage around. Here is GPT-3.5-Turbo:
| Position of the answer passage (of 20) | Accuracy |
|---|---|
| 1st | 75.8% |
| 5th | 57.2% |
| 10th | 53.8% |
| 15th | 55.4% |
| 20th | 63.2% |
With no passages at all, the same model scored 56.1%, so burying the answer mid-list did worse than giving it nothing. More material also bought little: going from 20 to 50 retrieved documents added only about 1.5 points for GPT-3.5-Turbo and about 1 point for Claude-1.3.
For a writer, the lesson is simple. The 2,000-word "master prompt" stuffed with every tip you've ever read is more likely to bury your instructions than to help. Pick what this one chapter needs.
Where should the question and the source material go?
Put long source material first and the actual task last. Anthropic's long-context guidance says to place long documents near the top of the prompt, above the query, and reports that putting the query at the end can improve response quality by up to 30 percent in their tests, especially with multi-document inputs of 20,000 tokens or more. Google's Gemini long-context guide gives the same advice: when the context is long, put the question after everything else.
That matches the Lost in the Middle curve, where the start and end of the input are used best. A layout that works for chapter drafting:
- Pasted source or research notes first, each piece labeled.
- Context: reader, book, tone.
- A short sample of the voice you want.
- Three to five must-follow rules.
- The task, in one sentence, last.
Anthropic also recommends separating these parts with XML-style tags and keeping tag names consistent, which is what the prompt above does. Labeled sections are plain text, so you can paste the same layout into other chat tools and test whether it helps there. The same order works for edits: paste your draft first, then say how much to change, as described in how to get AI to edit without rewriting everything.
How many examples should a prompt include?
One short sample is fine for a simple pattern; use three to five when the pattern is hard to describe. Anthropic's docs recommend 3 to 5 examples for best results, and say they should be relevant to your real task, varied enough that the model doesn't copy accidental quirks, and wrapped in tags so they aren't mistaken for instructions.
For voice, a sample of two or three sentences in your own style is easier for a model to imitate than adjectives like "friendly but authoritative", and it costs roughly 30 to 60 words. If the output still sounds generic after that, see how to make AI writing sound less like AI. For a full chapter structure, two or three example chapters can teach more than a list of rules, which is the approach in few-shot examples for consistent eBook chapters.
What should I do when my prompt keeps growing?
Split the job into steps instead of making the prompt longer. If you can't state the task in one sentence, you have two tasks. A plan prompt, a draft prompt per chapter and an edit prompt each stay short, and you can check each output before it feeds the next one. That pattern is prompt chaining, and prompt chaining for beginners walks through it.
The split is how Interactive eBook Builder is organized: Agent A Spec Builder produces the Chapter Plan, Agent B Drafter writes from it, and Agent C Polisher handles clarity, voice, pacing and layout, so no single set of instructions has to carry the whole book.
When you're troubleshooting a prompt that has already grown, these are my working heuristics rather than measured results:
| Symptom | Likely cause | Try this |
|---|---|---|
| The output ignores one of your rules | Too many rules, or the rule sits mid-prompt | Cut to 3 to 5 rules and move the key one next to the task |
| The output repeats your notes word for word | Notes pasted with no instruction on how to use them | Say "use only as source material; write in your own words" |
| The output is generic | No voice sample or reader details | Add a 2 to 3 sentence sample and one line on the reader |
| The answer drifts off the task | Task buried among background | Move the task to the last line |
How do I know which part of a prompt to cut?
Remove one block, run it again, and compare: if the output is no worse, leave the block out. Because outputs vary from run to run, a single run proves very little. Here is the routine:
- Save your current prompt and three outputs from it.
- Delete one block: the example, a single rule, or the context line.
- Run the shortened prompt three times on the same chapter.
- Compare against the saved outputs on what you care about: voice, length, and whether facts match your notes.
- Keep the cut if nothing got worse; restore the block if it did. Then test the next block.
Start with the likeliest passengers: repeated rules, vague adjectives and politeness lines. Leave the task, your notes and the voice sample for last. The test is quick when the whole prompt is visible, and in Interactive eBook Builder the agents' instructions are editable, so trimming a line and rerunning takes a minute instead of a rewrite.
Before you hit send
- Can you state the task in one sentence? If not, split it.
- Is each block answering something the model can't guess?
- Is long source material at the top and the task at the end?
- Are there 3 to 5 rules, not 12?
- Have you cut one block and compared the results?
Match prompt length to the task, put the task last, and cut anything that doesn't change the result.