AI Guides › Workbench
By Nigel Guy · 8 min read
The usual way to use AI for business ideas is to ask it for business ideas. It obliges, instantly, with a tidy list that sounds plausible and rests on nothing: no named customer, no evidence anyone pays to fix the problem. You feel productive and you have learned nothing you could not have guessed.
The rule: never ask AI for ideas; ask it for evidence of people already being annoyed, then try to kill whatever survives.
The kit has three prompts in a fixed order. Two run in a research tool that shows its sources (Perplexity is the example here), and one runs in a general chat assistant (ChatGPT) that argues against you. Each prompt hands its output to the next, and nothing gets built until the third has failed to knock the idea down.
| Step | Tool | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|---|
| 1. Complaint hunt | Perplexity | Searches the open web for recurring, specific complaints in a niche | Free tier has limited Pro Searches per day; Pro price varies, so check the £ price at checkout | Finding pain in the customer's own words | It only sees public pages, not private groups or what people say aloud |
| 2. Opportunity check | Perplexity | Looks for existing solutions, what they cost and where they fall short | Same as above | Telling "nobody solves this" from "nobody solves this well" | Pricing pages change; verify any figure it quotes |
| 3. Pressure test | ChatGPT | Argues against the idea, names the cheapest test | Free: $0 (£0); Go: $8 (about £6 at time of writing); Plus: $20 (about £15 at time of writing); Pro: from $100 (about £75 at time of writing). Check the £ price at checkout | Finding the weak assumption before you spend money | It has no new evidence; it can only reason from what you paste in |
You can use any research tool that cites sources and any chat assistant that follows instructions. The order matters more than the brand.
In Perplexity, use Pro Search or Research if you have it. The prompt asks for quoted, linked complaints and forbids it from proposing solutions, which is what stops it drifting into idea theatre.
You are a market researcher who reads customer complaints for a living.
Niche: [NICHE_OR_CUSTOMER_TYPE]
Region: [COUNTRY_OR_REGION]
Sources to favour: [E.G. FORUMS, REVIEW SITES, REDDIT, TRADE COMMUNITIES]
Goal: find what people in this niche repeatedly complain about, in their own words. Do not suggest products, ideas or solutions.
Steps:
1. Search the sources above for complaints, frustrations and "is there a tool for..." questions from the last [TIME_WINDOW, E.G. 24 MONTHS].
2. Group what you find into at most 8 recurring problems.
3. For each problem give: a one-line label, 2-3 short verbatim quotes with the source link and date, and how often it appears (many / some / a few, with the reason for your judgement).
4. Mark any problem you only found in a single source as "single source".
Rules:
- Only quote text you actually found. If you cannot find a quote, say so.
- Do not invent numbers or paraphrase and present it as a quote.
- If the niche is too broad to search well, ask me one narrowing question before you start.
Before answering, check that every quote has a link and that you have not suggested any solution.
Fill in the niche, region, source types and time window. Paste the result into your document and circle the one or two problems that appear most often and sound most expensive in people's own words.
Still in Perplexity, start a new thread so the first answer does not bias it, and paste only the problem you circled.
You are a competitive analyst. I want to know whether this problem is already well solved.
Problem: [PROBLEM_LABEL_AND_QUOTES_FROM_STEP_1]
Customer: [NICHE_OR_CUSTOMER_TYPE]
Region: [COUNTRY_OR_REGION]
Steps:
1. List the existing ways people solve this today: paid products, free tools, services, and workarounds such as spreadsheets or doing it by hand.
2. For each, give the vendor's own pricing page or description (link), what it charges if stated, and the complaints reviewers make about it (with links).
3. Say which complaints from the problem above remain unaddressed by every option.
4. State what evidence shows people pay to fix this (for example paid tools with real reviews, agencies, job listings). If you find none, say "no payment evidence found".
Output: a table with columns Option / Price (as stated, with date seen) / What it solves / Reviewer complaints / Source.
Rules: label anything you could not confirm as "unverified". Do not estimate market size. Ask me for missing details instead of guessing.
Before answering, check that each price has a source link and that you have separated stated facts from your inference.
The result you want is a specific unmet complaint with evidence of payment around it. A crowded market with a shared, unaddressed complaint is a gap. An empty market with no payment evidence is usually a sign that nobody cares enough.
Move to ChatGPT. Paste the circled problem, the competitor table and your own one-line idea. The prompt makes the model attack rather than flatter.
You are a sceptical advisor whose job is to find the weakest assumption in my plan before I spend money.
My idea in one line: [IDEA]
Target customer: [CUSTOMER]
Evidence so far (pasted from my research): [PROBLEM_QUOTES_AND_COMPETITOR_TABLE]
My budget and time for a first test: [BUDGET_IN_POUNDS_AND_WEEKS]
Steps:
1. List the assumptions the idea depends on (customer, problem, willingness to pay, how I reach them, why existing options fail).
2. Rank them by how badly the idea fails if each is wrong.
3. For the top three, give the strongest argument against me, using only the evidence above or clearly labelled general reasoning.
4. Propose the cheapest test for the single riskiest assumption, with a pass or fail threshold I set before running it.
5. Give a verdict: build, test first, or drop, and say what would change it.
Rules: do not praise the idea. Do not invent market data. If my evidence is too thin to judge an assumption, say what I need to find out. Ask me questions if anything above is missing.
Before answering, check that the verdict follows from the evidence and that the test could actually fail.
Run the test it proposes, not a bigger one. A landing page, a short interview round or a pre-sale each need little money. If you want the model's view on a draft test, paste the draft back and ask what result would embarrass you.
Hypothetical: you are curious about small dog-grooming salons in the UK. Step 1 returns complaints about no-shows, handwritten diaries and deposit chasing, with the no-show quotes appearing across several sources. Step 2 finds existing booking tools, mostly generic, and reviewers complain that deposits are awkward to take for short appointments. Step 3 identifies "salons will pay to stop no-shows" as the weak assumption and proposes asking ten salon owners to pre-pay a small monthly fee, passing at three yeses. Nothing has been built, and the idea is either alive for a reason or dead cheaply.