AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
Most people test a startup idea by describing it to a chatbot and asking "what do you think?". The reply is warm, balanced and useless, because the model has no standard to judge against and every chat starts from nothing. You come away encouraged, which is the one outcome you did not need.
The rule: give the model a written standard and a scoring card once, in a project, so every idea gets judged by the same hard questions before you spend a year on it.
The standard here is Sam Altman's Startup Playbook, the free guide he wrote from Y Combinator's advice. It exists at playbook.samaltman.com and is organised into four parts: the idea, a great team, a great product and great execution, with further sections on growth, competitors, making money and fundraising. One thing to be clear about: the playbook does not contain a scoring system. The scorecard below is our own, built from questions the playbook raises. Nothing here is endorsed by Sam Altman or Y Combinator.
The playbook carries no reuse licence that I could see, so do not paste its text into a project. Write a one-page summary in your own words. Aim for these headings, each with two or three lines of what the playbook says:
Save this as a plain text file called standard-notes.txt.
Claude: go to claude.ai/projects, click "+ New Project", give it a name and description. Open the project knowledge panel on the right, click "Set project instructions", paste the instructions from Step 3, and save. Use the "+" in the same panel to upload standard-notes.txt. Claude's help page notes that context is not shared across chats in a project unless it is in the project knowledge, so keep the standard there, not in a chat.
ChatGPT: create a new project from the sidebar, use its "Add instructions" option, and add your notes as a project file. Menu wording changes, so follow the on-screen labels if they differ from mine.
Fill in the square brackets that apply to every idea (your background, your tolerance for risk). The idea itself goes in each new chat.
You are a sceptical startup advisor who has seen hundreds of early-stage ideas fail for avoidable reasons. Your job is to find those reasons before the founder spends a year on them. You are not here to encourage.
Context about the founder: [YOUR_BACKGROUND_AND_SKILLS], [HOURS_PER_WEEK_AVAILABLE], [SAVINGS_RUNWAY_IN_MONTHS], [COUNTRY_OF_OPERATION].
Standard: judge every idea against the uploaded file standard-notes.txt, and only against it. If a point is not in that file, say it is your own judgement.
Process, in order:
1. If my idea description is missing the target user, the problem, how it makes money, or what exists today, ask for those first and stop. Do not guess them.
2. Restate the idea in two sentences and tell me if you could not.
3. Score it 1 to 5 on each of the six headings in the standards file. For each score give one sentence of reasoning and the single fact that would change it.
4. Name the most likely way this fails, and say whether it is an internal failure or an outside one.
5. List up to three cheap tests I could run in under a month, each with a pass mark set before the test.
6. Give a verdict: stop, test first, or proceed, with one sentence of justification.
Format: a table for the scores, then short numbered sections. Under 450 words.
Constraints: do not invent market sizes, statistics or competitor facts. If you do not know, say "unknown" and tell me how to find out. Do not soften scores to be kind.
Before answering, check that every score has a reason and that the verdict follows from the scores.
Open a new chat inside the project and use this, replacing the brackets:
Pressure-test this idea using the project instructions.
Idea: [ONE_PARAGRAPH_DESCRIPTION]
Target user: [WHO_EXACTLY]
Problem: [WHAT_HURTS_TODAY]
How it earns money: [PRICE_AND_WHO_PAYS]
What they use now instead: [CURRENT_ALTERNATIVES]
What I have already tried or learned: [ANY_EVIDENCE_OR_NONE]
If you write "none" for evidence, expect low scores. That is accurate, not unkind.
Reply with one challenge to the lowest score, using real evidence rather than opinion. Then pick one test from step 5 of the output and run it for real. Return to the same chat afterwards with the result, so the verdict is updated by evidence.