# The brainstorm-iterate loop, and one loop for four kinds of ask (/docs/prompting/the-brainstorm-iterate-loop) --- type: Procedure title: "The brainstorm-iterate loop, and one loop for four kinds of ask" description: "Concept 7 of the book's prompting guide, the single highest-leverage habit: load context, demand options, give explicit feedback, iterate, expand only the winner, then grade and re-iterate until the score plateaus; outline before drafting; research before either; and the dial that loosens the spec for options and tightens it for execution." status: stable order: 5 generated: { by: "process:claude-code", at: 2026-09-03T08:15:00Z } verified: - { by: "human:rrizwan98", at: 2026-09-03T08:20:00Z } sources: - id: af-03 title: "AI Prompting 2026: 13 Concepts, 80% of Real Use. The AI Agent Factory, Panaversity, 2026" resource: https://agentfactory.panaversity.org/docs/ai-prompting-2026 ksor: audience: [public] owner: human:rrizwan98 approval: { by: "human:rrizwan98", at: 2026-09-03T08:20:00Z } trust_tier: human-reviewed build_id: sha256:4d4bff09bd0a1dcff0a0a50f20f7ef514fc471e2b1f9e81e0e7b6150b93c0793 dirty: true --- This is the single highest-leverage habit in the guide. If you skip every other page, do not skip this one. [^af-03] When AI was trained on the internet, most of the internet was common ideas, not creative ones. So the average AI response on a creative question is also common. "Ways to exercise at home": squats, push-ups, planks. Not wrong. Just average. The way around this is not a magic prompt. It is a loop. [^af-03] ## The recipe 1. **Give all relevant context up front.** Not "ways to exercise"; "ways to exercise given that I have stairs in my home, a bad knee, and I cannot stick to plans for more than three days." 2. **Ask for 3 to 5 options, not one.** Forcing alternatives pushes the model past its first instinct. 3. **Give explicit feedback.** "I don't like option 1, it's too passive. I do like the stair-climbing idea but want it shorter. I forgot to mention my knee gets worse on impact." 4. **Ask for 3 to 5 new options informed by the feedback.** 5. **Iterate until you have one or two you genuinely like.** 6. **Then, and only then, ask AI to flesh out the chosen option in detail.** [^af-03] The book's worked example, debt payoff: [^af-03] ```text I have $8,000 in credit card debt at 19% APR, $4,000 in student loans at 5%, and $1,200 in a retail card at 24%. I have $700/month free after expenses. I just learned I'll get $450 in cash from a tax refund. Risk tolerance: low. I sleep badly when I see big balances. Give me 5 different repayment strategies, each with a one-line rationale. Don't expand any of them yet. ``` Then, after reading the five options: ```text Reject option 2 (avalanche by interest rate alone): I want psychological wins early. Reject option 4: I won't open new accounts. I like option 1 (snowball with the retail card first) but I'd want to fold the $450 in. Give me 5 new options that combine snowball-style wins with smart use of that lump sum. ``` You are not waiting for the AI to read your mind. You are showing your taste; the AI reshapes the option space around it. After two or three rounds you have one option that feels exactly right. Then ask for the full plan. [^af-03] **Grading is one way to iterate; diagnosis is the other**, and for a weak first result it is the stronger one. Instead of a score, read the output and name which part of your request failed it: the audience was ignored, the length constraint broke, the tone drifted. Change only that part and resend. A score tells you how far off you are; naming the failed part tells the model exactly what to move. When a draft misses badly, diagnose first and grade later. [^af-03] ## Outline before drafting The same loop works for writing, where it has its own name: [^af-03] ```text - Iteration 1: ask for 3 outline options for a post on X. - Iteration 2: pick one outline, ask AI to critique it and grade it out of 10. Note what scored below 9. - Iteration 3: revise the outline based on the critique, then ask AI to expand each heading into 3 to 5 bullets. - Iteration 4: critique the bullets, grade them out of 10, fix the ones below 9. - Iteration 5: only now ask for the full draft. - Iteration 6: critique the draft, grade it out of 10, ask for the changes that would raise the score the most, ranked by impact. Repeat until the score plateaus around 9.5 or higher. That is your stopping signal, not "the AI says it is done." ``` Why this works: editing one word in an outline can change the direction of the whole article; editing one word in a final draft changes one word. Almost all the leverage in writing happens at the outline level. AI generates word by word from the start, so unless you force structure first, it cannot see the whole shape. **Don't skip steps.** AI's first draft of anything is slop: looks polished, says little. Ten or twelve minutes of structural work before any drafting, then several rounds of grade-and-fix, turns a forgettable post into one that lands; the total is rarely more than forty-five minutes for a 600-word piece. [^af-03] ### The worked writing example A team lead wants a 600-word post titled "Why our small AI team is shipping faster than the big team across the hall." [^af-03] Round 1, research first: ```text I'm writing a 600-word post arguing that small AI-augmented teams ship faster than larger non-AI teams. Don't write yet. First, give me the 5 strongest research-backed arguments and the 3 strongest counter-arguments. One sentence each. ``` Round 2, three outlines: ```text Now produce 3 different outline options for the post. Each outline should have 4-6 headings. They should differ in structure: one narrative, one analytical, one contrarian. One line per heading. ``` Round 3, pick one and add an analogy: ```text I'll go with outline 2 (analytical). I want to weave in a Pixar analogy: how the original Toy Story team was small and faster than the giant Disney studio because of new tools. Add this as a recurring example, not its own section. Revise outline 2. ``` Round 4, expand to bullets: ```text Now expand each heading into 3-5 bullets. Telegraphic style, not prose. ``` Round 5, grade and fix the bullets: ```text Critique each bullet and grade it out of 10 with a one-sentence justification. List the bullets scoring below 9. For each one, suggest the change that would raise the score the most. ``` Only now does the lead ask for the full draft, and then keeps grading and re-iterating until the score plateaus around 9.5 or higher. The output reads like the lead wrote it, because every load-bearing decision was the lead's. [^af-03] **Scope the territory before drafting.** Round 1 looks small but does heavy work. Most people skip it and ask for the draft directly, which is why their drafts feel thin: they are built on whatever ideas the model surfaces first, not on the actual landscape of the topic. This generalizes far past writing. Before any substantial decision, plan or analysis, ask the AI to map what is known before asking it to produce what is needed. Competitive landscape before product naming. Prior research before a strategy memo. Existing approaches before designing a new one. [^af-03] ## The loop is domain-agnostic Load context, demand options, give explicit feedback, demand new options, iterate, expand, then grade and re-iterate until the score plateaus. If you find yourself accepting the AI's first answer, or stopping the moment something looks good enough, you have skipped the loop. [^af-03] | Decision or task | What "context" looks like | What "options with feedback" looks like | | --- | --- | --- | | Planning a 4-day trip | Constraints (budget, dates, who's going, what they hate) | 5 itinerary skeletons; reject two; iterate the rest | | Naming a product | What it does, who buys it, what it must NOT sound like | 10 names; pick 3 you like, ask for variants on those | | Writing a difficult email | The recipient, the relationship, the desired outcome | 3 different tones; pick one, refine its specifics | | Choosing a contractor | Three quotes, three reference notes, your priorities | Side-by-side scoring; ask for the strongest counter to your favourite | | Picking a learning path | Current skills, time available, end goal | 3 curriculum shapes; pick one, expand to weekly milestones | | Designing a logo brief | Brand values, audience, examples you like | 5 mood-board directions; pick one, ask for 5 variants in that lane | In every row, once you have a concrete candidate, the grading move applies: score it out of 10 against the criteria that matter for that task, then iterate. The criteria change; the move does not. [^af-03] **Two stopping rules, not one.** The 9.5 plateau is the standard for work you own and will keep using: a portfolio piece, a template, a page like this one. For a task you are delivering once and moving on from, stop earlier: the round where another prompt produces only marginal change, and your own two-minute manual edit would finish the job faster, is the round to take the wheel. [^af-03] ## One loop, four kinds of ask Not every request is a brainstorm, and the most common way good prompters go wrong is running the right loop with the wrong grip. A marketer wants campaign concepts. His request is a model brief: the channel mix, the exact tagline, a required three-part structure, the tone, a 40-word cap per concept, five concepts please. The AI returns five near-identical rewordings of one idea; asked for twenty, he gets twenty neighbours of the same idea. Nothing is broken. His constraints pinned the structure, the tagline, the tone and the length, leaving the model one small neighbourhood of answers to sample from. [^af-03] **The fix is not more words. It is fewer constraints, for now.** Ask for five genuinely different directions with only the true constraints attached: the product, the audience, what it must never sound like. When one direction wins, bring the full specification back and have that direction executed precisely. Precision is the right setting for execution and the wrong setting for generating options. [^af-03] | Kind of ask | You want | Tighten | Loosen | The failure if you get it backwards | | --- | --- | --- | --- | --- | | **Brainstorming** | genuinely different directions | the problem, the audience, what to avoid | structure, tone, format, length | five rewordings of one idea | | **Research** | the landscape mapped | scope, sources, what counts as evidence | the answer; never say what you expect to find | a map of your assumptions, not the territory | | **Drafting** | one thing, executed | everything: voice, length, structure, facts, format | almost nothing | a fluent draft of the wrong thing | | **Analysis** | what the data actually says | the data, the definitions, the exact question | the conclusion; never specify it | an analysis that flatters the ask | This looks like a contradiction of the context page and is not. Context about your **situation** is never the problem; the model cannot know too much about the budget, the audience or the bad knee. The dial controls how tightly you pin **the shape of the answer**. Load the situation fully every time; specify the output loosely while diverging and precisely once converging. The four kinds chain, and the writing example ran the chain: research, then brainstorming, then drafting, then analysis. When a task feels stuck, the commonest reason is sitting in the wrong column. [^af-03] Continue to [beyond text](beyond-text.md). [^af-03]: AI Prompting 2026: 13 Concepts, 80% of Real Use, Part 2, Concept 7 "The brainstorm-iterate loop" and "One loop, four kinds of ask". The AI Agent Factory, Panaversity, 2026.