A seven-map framework for deciding what to build with AI, the prompts to run it across different models, and a free tool that tells you to build, reshape or park
A founder came to us with a one-line brief: "An AI agent that does all of a user's computer tasks."
The technology was real. AI agents that can see a screen, click and type became genuinely usable this year, and their prototype worked. Srikanth read the brief and asked the question we now start every product engagement with:
"Who is it not for?"
"Well," the founder said, "it's for everyone who uses a computer."
That's the most common answer we hear, and the most expensive. A product for everyone has no first customer, no message that lands, no channel that works, and no defence when an AI company ships the same capability as a free feature next month.
This post covers the framework we use to answer that question before anyone writes code, the prompts to run it, how to use different AI models for different jobs, and a free tool that tells you whether to build, reshape or park your idea.
Building got cheap. Deciding didn't.
Two years ago, the hard part of a startup was building the product. Today, an AI coding agent can produce a working prototype in a weekend. The cost of how has collapsed.
The cost of what and for whom hasn't. When anyone can build anything, the decision about what to build is the main thing that separates products.
The trap is that people now use AI for that decision too, and they use it badly. They ask "Is this a good idea?" or "Who's my target customer?" and get back something fluent, confident and generic.
Same idea, two questions. The generic question gets the answer every founder gets. Illustrative outputs.
AI models are trained toward the most probable answer, and the most probable target customer for any productivity tool is "busy professionals and small businesses who value efficiency." That line has appeared in thousands of pitch decks. It will appear in yours unless you ask a better question.
You can sell anything if you know what to focus on
I've long believed that if you know how to sell, you can sell almost anything. But selling well isn't about being persuasive. It's about knowing what to focus on: which pain, which person, which moment, which frame.
The best recent example is Clicky. On paper, it's a screen-aware AI agent: it sees your screen, listens to your voice and can act on your Mac. That's the same category as our founder's brief, and as dozens of other products.
But it isn't sold as "an agent that does your tasks." Its Product Hunt listing calls it an "AI buddy next to your cursor" that sees, guides and helps. It's framed as company: help that sits beside you, right where you're stuck. It names a modern problem everyone feels (too many apps, too much to figure out alone) and offers a modern answer.
Same capability, completely different frame, and the frame is what people relate to and pay for.
People don't buy the capability. They buy the frame.
The lesson: positioning is a product decision, not a marketing afterthought. The frame decides who you build for, which features matter, what you charge and where your first users come from.
The negative-space framework
In design, negative space is the empty area around a subject, and it's what gives the subject its shape. Products work the same way: an idea gets its shape from what it deliberately isn't.
Our framework asks seven questions. Each comes with a negative-space question, the one most founders skip.
Seven maps. The negative-space question is where focus comes from.
1. Pain: what hurts, how often, and what does it cost today? Write the pain in the customer's own words. Then measure how often it happens, how badly it hurts and what they already spend on a workaround. If people already pay, in money, time or a hired person, there's a budget waiting. Negative space: what pain will you ignore, even if users ask for it?
2. Person: exactly who, in which moment? A useful definition has a role, a context and a trigger moment. "Accountants" is a role. "Solo chartered accountant (CA) practices with 40 to 150 GST clients" adds context. "In the ten days before each filing deadline" adds the moment the pain peaks. Negative space: who are you deliberately NOT building for, and why? This is the most important question here. Name at least two groups you'll turn away: enterprises (nine-month security reviews), tinkerers (they'll build their own), people who want full autopilot (one wrong click and they leave). Every "no" makes your "yes" more believable.
3. Promise: what do you promise, in one line? The structure is simple: we help [who] [do what] without [the pain]. The craft is the frame: a relationship, a rescue, an identity. Write twenty versions and cross out every one a competitor could also say. Negative space: what will you NOT do, so the promise stays sharp?
4. Proof: why now, why you, and what evidence? What changed to make this possible now? What access, data or insight do you have that others don't? ("We use AI" isn't an answer.) And where are you on the evidence ladder: opinion, interviews, waitlist, letters of intent, paying customers? Negative space: what evidence would prove you wrong, and have you looked for it?
5. Path: how do the first 100 users find you? Name a specific place your customers already gather: a community, an association, a marketplace, a partner. Negative space: which channels will you NOT use? If paid ads are your only channel, your unit economics will decide your fate before your product does.
6. Price: who pays, how much, for what unit of value? Choose a pricing unit that grows with the value you create: per seat, per document, per filing. Note whether the payer and the user are different people. Negative space: below what price will you NOT sell?
7. Durability: why does this survive when AI models get better? Ask bluntly: if an AI company shipped your core feature as a free update next month, what would be left? Durable AI products own a workflow, proprietary data, customer trust or distribution. Negative space: when will you stop? Write kill criteria as a number and a date: "If fewer than 5 of 10 paid pilots renew by March, we stop."
Use AI as a co-founder, not an oracle
Most people open one chat with one model and ask it to judge their idea. That gives you one opinion, shaped by one model's training, and inclined to agree with you.
Instead, staff the room. Different models are good at different jobs.
Five roles. Two or three different models can cover all of them.
•The Researcher: a model with web search, for verifiable facts: competitors, pricing, real customer complaints. Always check the sources.
•The Skeptic: your strongest reasoning model, for pre-mortems and attacking your "not for" list.
•The Rehearsal Partner: any capable model playing a customer, to practise interviews. Never treat its answers as evidence; simulated customers aren't customers.
•The Wordsmith: a fast, inexpensive model run many times, for volume: twenty taglines, ten names. Ask it to write a random string first and build from it, so it doesn't repeat itself.
•The Second Opinion: a model from a different company. Where two vendors disagree is exactly where you should think hardest.
Four rules make all of them more useful: never ask "Is my idea good?", require at least one disagreement, demand numbers as ranges, and keep generating separate from judging.
The prompts
Start every planning conversation with an idea brief. It's the single biggest upgrade you can make, because it stops the model falling back on the average startup.
IDEA BRIEF (read before anything else)
- What I'm building, in one line:
- Exactly who it's for (role + context + trigger moment):
- Who it's NOT for, and why:
- The pain, in their words:
- What they do about it today:
- What I already know (evidence, numbers):
- My constraints (budget, team, time, geography):
Rules: no generic advice. Disagree with me at least once.
Give numbers as ranges, and label guesses as guesses.
Then use these, matched to the model roles above:
Person (Skeptic):
Name 3 groups I should refuse to sell this to. For each, explain why
they'd churn, cost too much to serve, or damage the product. Then
describe the narrow group that's left, and the moment its pain peaks.
Promise (Wordsmith, run several times):
First write a random 16-character string. Using only that string to
drive your choices, write 20 one-line promises, each with a different
emotional frame (rescue, relationship, identity, relief, control...).
Then cross out every line a competitor could also say.
Proof (Skeptic):
List 5 pieces of evidence that would prove this idea wrong, and the
cheapest way to look for each within 14 days. Then design one test of
my riskiest assumption using real money or commitments, not opinions.
Durability (Skeptic):
It's two years from now and this product has failed. Write the post-
mortem, including one scenario where an AI company shipped our core
feature for free. What would we need to own to survive each scenario?
Second opinion (a different vendor's model):
Here's another AI model's assessment of my idea. Don't be polite. What
did it get wrong or miss, and where would you reach a different verdict?
[paste the other model's answer]
Verdict: map your idea, get a decision
Prompts are powerful, but they mean pasting an unreleased idea into someone else's model, and they give you prose rather than a decision. So we built Verdict.
You fill in the seven maps, and it gives you:
•a metric map scoring each map from 0 to 5,
•a verdict: Build, Reshape or Park,
•your three weakest areas, each with a specific next step,
•a positioning draft built from your answers,
•a 14-day plan with no code until day ten.
Verdict starts where most planning skips: who it's not for.
Each map pairs its questions with the negative-space question, and the "not for" fields are clearly marked:
The Person map: who exactly, who not, and whether you could name real people who fit.
Scoring combines your own ratings with checks on how specific your answers are. Vague words like "everyone" or "businesses" lower your score, and so does "we use AI" as your advantage. It's deliberately strict, because the market will be.
Like our compliance tool Preflight, Verdict has no AI and no server. It runs entirely in your browser and blocks every outgoing request, so your idea never leaves the page.
Try it: verdict-two-wheat.vercel.app, or load the two examples from this post: the generic idea and the same idea, reshaped.
The same idea, scored twice
We put the founder's original brief into Verdict exactly as it came to us: an agent for everyone, marketed through social media and ads, with no evidence beyond enthusiasm and no answer to "what if an AI company ships this?"
The original brief: 1.4 out of 5, every area in the red. Verdict: park it.
The technology wasn't the problem; nothing around it had been decided. Verdict's top suggested changes pointed straight at the gaps:
Verdict's three priority changes for the original brief.
Then we spent two weeks on the maps with the founder, not on code. We kept the capability and changed everything around it:
•Person: solo and two-partner CA practices in India with 40 to 150 GST clients, in the ten days before each filing deadline.
•Not for: large firms with in-house tech teams, businesses filing their own returns, and anyone wanting full autopilot without review.
•Promise: "Your filing-week co-pilot: it sits next to your cursor, sees the portal you see, and does the copy-paste-reconcile while you review."
•Won't do: file anything without the CA's approval, give tax advice, or sell to enterprises in year one.
•Proof: the founder had run a CA practice and had 120 CA contacts, which turned into interviews, a waitlist and paid pilots.
•Path: professional WhatsApp groups and association meetups, then referrals.
•Durability: a reconciliation rulebook built from thousands of real filings, plus the trust of a buyer who will never hand a government portal to a general-purpose agent.
Same capability, reshaped: 4.3 out of 5, every area above the line. Verdict: build.
The capability didn't change. The decisions did. And on a sales call, nobody says "an AI agent that does your tasks." They say, "Someone finally understands my 2 a.m."
What to do this week
AI has made building the easy part. The advantage now belongs to people who decide well, and deciding well is a skill you can practise.
1.Write your "not for" list: at least two groups, with a reason for each.
2.Rewrite your one-liner in three frames and test them with five real customers.
3.Climb one rung on the evidence ladder. If you have opinions, get interviews; if you have interviews, ask for money.
4.Answer the durability question honestly.
5.Write kill criteria as a number and a date.
6.Run your idea through Verdict or the prompts above, with at least two different models, before writing more code.
The founder's agent starts paid pilots with forty CA practices next month. It still sees the screen, clicks and types. But now it knows exactly who it's for, and who it isn't.
Decide what to build before you build it
The cost of building has collapsed; the cost of building the wrong thing hasn't. AtliQ helps founders and product teams validate ideas, sharpen positioning and turn the right idea into a production-ready AI product, with the strategy, design and engineering under one roof.
Have an AI product idea you want pressure-tested? Talk to our product team before you write the first line of code.
This is part three of a series on building with AI: the wins, the near-misses, and the people in the room when they happened.
Sources
