AI Guides › Playbooks
By Nigel Guy · 7 min read
"I don't know how to build that" used to kill most ideas before they cost you anything. With an AI model writing the code, the copy and the landing page in an afternoon, that excuse has gone, so ideas now go straight from shower thought to half-built prototype. The trouble shows up three weekends later: you have a working thing, nobody waiting for it, and a monthly bill to keep it alive.
The rule: building got cheap, but deciding didn't. An idea gets real time only after it has passed six written questions, and each answer has to be evidence you collected, not a sentence you or a model made up.
The gate is a single page, or a single note, that you fill in before you open an editor or a builder tool. It has six rows. Each row has a question, what counts as an answer, what doesn't, and the signal that should stop you.
| # | Question | Counts as an answer | Doesn't count | Stop signal |
|---|---|---|---|---|
| 1 | Who has this problem this month, by name? | Two or three real people or businesses you could message today | "Small businesses", "creators", "anyone who…" | You can't name one person |
| 2 | What do they do about it now? | The current workaround: a spreadsheet, a person, a paid tool, or putting up with it | "Nothing exists for this" | They don't do anything because it doesn't bother them |
| 3 | What have they already given up to fix it? | Money spent, hours lost, a tool they pay for and complain about | "They said it sounds great" | Only compliments, no cost |
| 4 | What is your appetite? | A fixed time budget written down before you start, e.g. two weekends | "As long as it takes" | You can't name a number you'd stop at |
| 5 | Who keeps it running in six months? | A named owner, the running costs (hosting, API usage, subscriptions) and what happens when a dependency changes | "It'll basically run itself" | The upkeep would land on you and you don't want it |
| 6 | What stops if you start this? | The specific project, client work or rest that loses the hours | "I'll fit it in" | The thing that stops matters more |
The gate is not a scoring system that tells you to build. It tells you whether you have enough evidence to spend real time. The GOV.UK Service Manual says the same about its discovery phase: stopping at the end of discovery isn't a failure when the research says stopping is the right call.
The trap is letting the same tool that makes building cheap also fill in the gate. Ask a model "who would want this?" and it will produce three plausible personas, a market-size paragraph and a confident pricing tier. None of it is evidence. It's confidence theatre: the right shape, nothing real inside.
Building first "just to see" is the same trap: a finished prototype makes people polite, and politeness reads like demand. Use AI on the gate in one way only: as an interrogator that pushes back on your answers. The prompt below does that.
You are a sceptical but fair advisor reviewing whether an idea deserves real time before anything is built. You are not here to encourage me or to generate ideas.
My idea, in one sentence: [IDEA_IN_ONE_SENTENCE]
My answers to the six gate questions:
1. Named people who have this problem this month: [NAMED_PEOPLE]
2. What they do about it now: [CURRENT_WORKAROUND]
3. What they have already given up to fix it (money, time, tools): [EVIDENCE_OF_COST]
4. My appetite (fixed time budget): [TIME_BUDGET]
5. Who keeps it running in six months, and the monthly cost: [OWNER_AND_RUNNING_COST]
6. What stops if I start this: [WHAT_LOSES_OUT]
Work through these steps in order:
1. For each answer, label it EVIDENCE (something observed, paid or said about a past event), ASSUMPTION (my belief) or MISSING. Quote the part of my answer you are judging.
2. For anything labelled ASSUMPTION or MISSING, write one concrete action I could take this week to replace it with evidence, such as a specific question to ask a named person about the last time the problem happened.
3. Mark each row PASS, WEAK or FAIL, with one line of reasoning.
4. Give a verdict: GO (within the stated appetite), NOT YET (list what to collect first) or DECLINE (say why).
Format: a table with columns Row, Label, Rating, Reason, Next action, then the verdict on its own line.
Rules:
- Do not invent customers, numbers, market sizes or quotes. If you need a fact I haven't given, ask me for it instead of estimating.
- If any placeholder above is still in square brackets or empty, stop and ask me for it before reviewing.
- Treat "people said it sounds great" as an assumption, not evidence.
- Do not suggest features or improvements to the idea.
Before you answer, check: did you label every row, did you avoid adding facts I didn't supply, and does the verdict match the ratings?
Fill in: your one-sentence idea and your six answers, in your own words, including the gaps.
Suppose you run a small bookkeeping practice and want to build an AI tool that turns clients' photographed receipts into categorised expense lines.
Verdict: not yet. Before building, check whether the accounting software you already use has a receipt-capture feature you aren't using, and price storage and API costs.