AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people with an AI project idea do one of two things. They wait until the idea is big enough to be worth the effort, or they ask for the whole thing at once: the app, the accounts, the dashboard, the lot. Both feel sensible, and both end the same way, with nothing built and a tidy sense that the idea needs more thought.
The idea that feels too small to bother with is usually the one that fits in a single sitting. That is its advantage, not its flaw.
The rule: build the smallest thing that does one job for one person, finish it today, and only then decide what it should grow into.
A small build gets finished, and a finished thing teaches you what a planned thing cannot. You learn what you actually use, what you ignored, and what you assumed wrongly. A big plan gives you none of that until the end.
There is a cost argument too. At time of writing, Anthropic's pricing page lists Claude's Free plan at $0 and says it includes Projects (up to five) and Artifacts, with Pro at $20 a month (about £15 at time of writing; check the £ price at checkout). A one-screen tool does not need more than the free tier to test whether it is worth having. Plans and limits change, so check the pricing page before you rely on this.
The mechanism is a card you fill in before you open a chat. It has five lines, and it has to fit on one screen.
| Line | What you write | Test |
|---|---|---|
| Person | The one person who will use it (often you) | Can you name them? |
| Job | One sentence: "When X happens, this does Y" | Does it contain "and"? If so, cut |
| Input | What you will type or paste in | Do you have a real example to hand? |
| Output | What a good result looks like, in your words | Could you spot a bad one? |
| Done when | A check you can do in five minutes | Does it use a real example, not an imagined one? |
If you cannot fill a line, that is the finding. Fix the card, not the prompt.
Paste this into a new chat once the card is filled in.
You are a practical helper who builds very small, single-purpose tools with
non-technical users.
Context: I want one small thing that does one job for one person. Here is my
card.
- Person: [WHO_USES_IT]
- Job: [ONE_SENTENCE_JOB]
- Input: [WHAT_I_WILL_PASTE_OR_TYPE]
- Output: [WHAT_A_GOOD_RESULT_LOOKS_LIKE]
- Done when: [FIVE_MINUTE_CHECK]
Here is one real example input: [REAL_EXAMPLE_INPUT]
Work in this order:
1. Check the card. If any line is missing, vague, or contains two jobs, ask me
up to three questions and wait. Do not guess.
2. Run the job once, by hand, on my example, and show the result.
3. Tell me whether the result meets my "Done when" check, and say plainly
where it falls short.
4. Only if I agree, propose the simplest way to repeat this: a saved set of
instructions first, a one-screen tool only if the job needs buttons or
fields.
5. Do not add features I did not ask for. List up to three possible
extensions at the end, one line each, and build none of them.
Constraints: keep it to one screen of output. Use British English. Do not
invent data, names or figures; mark anything you had to assume.
Before answering, check that you stayed within the single job on my card.
Fill in the five card lines and one real example.
Say Priya runs a small allotment society. Each month she rewrites the same reminder email from a pile of scribbled notes, and she thinks automating that is too trivial to bother with.
She runs it in a plain chat, and it passes. She then saves the instructions in a Project so next month starts from her notes alone. That is the whole build. Later she notices she keeps pasting the same rota, so she adds the rota file to the Project. The second step exists because the first was used, not because she planned it.