To turn a messy brain dump into a usable prompt, sort it into six buckets (task, reader, must-includes, rules, output format and unknowns), turn feelings into rules, and keep raw notes separate from instructions. You don't need to polish your thinking first. You need to label it.
Below is a five-step method, a full before-and-after example with word counts, and a way to hand the sorting to the AI itself.
Why doesn't a raw brain dump work as a prompt?
Because the model can't tell your instructions from your thinking out loud. A brain dump is full of real information, but it's unlabeled: a preference, a rule, a plan and a worry all look the same. Anthropic's prompting guide recommends being clear and direct and suggests showing your prompt to a colleague with minimal context to see whether they'd be confused. Try that with a dump and you'll find a person would ask the same five questions the AI should.
Here's a dump I wrote as an illustration (74 words):
ok so the book is about getting started with sourdough, mostly for people who failed before, my starter always died, i want it short, maybe 8 chapters?? i hate fancy words. oh and the part about feeding schedule is the most important, people get that wrong. include my story about the brick loaf. no weights in cups only grams because cups ruin everything. maybe a troubleshooting chapter at the end. friendly but not cutesy
All of the content is good. But: is "my starter always died" a fact about the reader, about the author, or both? Is "maybe 8 chapters??" a requirement? Is "most important" a placement rule or a length rule? A model will guess, and guesses regress to the average.
What are the five steps?
Sort, convert, separate, specify, and let it ask.
- Sort into buckets. Underline each fragment and drop it into one of six boxes (table below).
- Convert feelings to rules. "I hate fancy words" becomes "plain words, short sentences". "Cups ruin everything" becomes "grams only, never cups".
- Separate instructions from material. Your anecdotes, notes and source text go in a clearly marked block. Anthropic advises tagging distinct content and, for long documents, placing them above the question; OpenAI's example puts context after the instructions. Both agree on the principle: don't let material and instructions blur.
- Specify the output. Say what you want back: a numbered list, a table, three options, 600 words.
- List unknowns. Permit questions. See how to get AI to ask clarifying questions for a template.
| Bucket | What goes in it | From the sourdough dump |
|---|---|---|
| Task | The one thing you want now | Propose an 8-chapter outline (no drafting yet) |
| Reader | Who it's for, what they know | People whose starter died; want plain words |
| Must-include | Things that have to appear | Feeding schedule early; brick loaf story; troubleshooting chapter |
| Rules | Voice and constraints | Grams only; friendly, not cutesy; short sentences |
| Output | Shape of the answer | Numbered list, one-sentence purpose each |
| Unknowns | What you haven't decided | Chapter count (the dump said "maybe 8??") |
What does the sorted prompt look like?
It looks like the same ideas with labels, which makes it 111 words instead of 74. Here is the result of the five steps:
TASK: Propose an 8-chapter outline for a short beginner's ebook on sourdough. Do not write the chapters yet.
READER: People who have tried sourdough before and failed (their starter died). They want plain words, no jargon.
MUST INCLUDE:
- A chapter on the feeding schedule (the most common point of failure), placed early.
- My story about the "brick loaf" as the opening anecdote.
- A troubleshooting chapter at the end.
RULES: Measurements in grams only, never cups. Voice: friendly, not cutesy. Short sentences.
OUTPUT: A numbered list of 8 chapters, each with a one-sentence purpose.
UNKNOWNS: If something important is missing, ask me up to 3 questions before you outline.
| Measure | Brain dump | Sorted prompt |
|---|---|---|
| Words | 74 | 111 |
| Clear task stated | No | Yes (outline, not drafting) |
| Explicit rules | 0 (all implied) | 3 (grams, voice, sentence length) |
| Output format named | No | Yes |
| Placement of key chapter | "Most important" | "Placed early" |
| Permission to ask | No | Yes, up to 3 questions |
The word counts come from counting the two blocks above. The prompt isn't longer because it's wordier. It's longer because it adds decisions: placement, format, rules and a way out of ambiguity. If you're wondering whether that length is justified, our post on how long a prompt should be covers when extra words help and when they hurt.
Can I get the AI to do the sorting?
Yes, and it's a good use of the tool as long as you review the result. Paste your dump and ask:
Below is a messy brain dump. Do not do the task. Instead:
1. Sort it into: TASK, READER, MUST INCLUDE, RULES, OUTPUT, UNKNOWNS.
2. Turn any opinion or feeling into a concrete rule.
3. List anything contradictory or missing as questions.
4. Show the result as a prompt I can paste.
<dump>
[paste]
</dump>
Check three things in what comes back: did it drop anything you care about, did it invent a preference you never stated, and did it resolve a contradiction without telling you? A model that quietly picks one side of a contradiction hides the exact thing you needed to decide. This is the kind of review step that pays off most with ambiguous input, and the same caution appears in what prompt engineering means in plain English.
Three mistakes show up again and again when people sort their own notes. First, putting two tasks in one prompt ("outline it and also write chapter one"); split them. Second, leaving a number as a maybe ("maybe 8??") without saying whether it's a rule or a starting guess; write "8, unless you can justify a different number". Third, burying the most important instruction in the middle of the notes; put your top rule where you'd put it if you only had one line to say.
If you'd rather not do the sorting in a chat at all, Agent A Spec Builder in Interactive eBook Builder takes your setup and writes the Chapter Plan and the niche packs for you, which is the same sorting done as a step in a workflow.
What happens after the prompt is sorted?
You use it as the seed of a larger plan, not as the end of the work. Once the outline comes back and you've edited it, the next prompt is a chapter prompt that carries the outline forward; prompt structure for turning an outline into an ebook draft shows how.
The same sorting is built into book setup in Interactive eBook Builder. Its Configuration Hub asks for the decisions your dump contains (title, reader, preferences) as fields rather than prose, and the Seed Prompt then initializes the eBook's overall structure, Niche Spec and chapter structure from them. Doing the buckets by hand, as above, is the manual version of that same step, and it works in any chat window.
Before you hit send
- Is there exactly one task, and does it say what not to do yet?
- Is every preference rewritten as a rule you could check?
- Are notes and anecdotes in a separate, labeled block?
- Did you say what shape the output should take?
- Did you allow the AI to ask about gaps?
The takeaway: you don't need cleaner thoughts, you need labeled ones. Six buckets turn a ramble into a brief.