AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people worried about AI treat it as a contest of knowledge: the model has read everything, so what is left for you? So they spend their effort learning tools, collecting prompt lists and chasing the newest model name. That feels productive while it quietly throws away the one input the model cannot supply, which is what you know about your own field and how it fails.
The rule: the model supplies drafts; you supply context, the test for "good" and the final check. Hand over only what you can judge.
The mechanism is less mysterious than the fear. A model works from what is in the conversation. Anthropic's own prompting guidance tells you to think of Claude as "a brilliant but new employee who lacks context on your norms and workflows", and says that giving the reason behind an instruction helps it deliver more targeted results (Anthropic prompting best practices). Someone with ten years in a field knows the norms, the exceptions and what a wrong answer looks like. Someone with ten hours of tool tips does not, and cannot tell a confident error from a correct answer.
Three things sit on your side of the table, and none can be installed by a tool tutorial.
There is research pointing the same way, with a caveat. A 2023 field experiment by researchers from Harvard Business School, Wharton, MIT Sloan, Warwick and BCG gave 758 BCG consultants realistic tasks, some with GPT-4 assistance. The authors describe a "jagged technological frontier": on tasks inside it, AI users did better; on a task outside it, those using AI were less likely to reach the correct answer than those without (SSRN working paper, as summarised in search results; I could not open the full paper, so check it for the exact figures). The lesson is not "AI is great" or "AI is risky". It is that you need to know where the edge is in your own field, and only experience tells you.
Fill this in once per recurring task. It takes ten minutes and replaces a lot of vague prompting.
| Row | Question | Your answer (example) |
|---|---|---|
| Task | What exact job is this? | First draft of a reply to a tenant's repair complaint |
| Context only I hold | What would a new hire not know? | Our repairs policy, tone with this landlord, the last three disputes |
| What "good" looks like | How would I mark this? | Correct timescales, no admissions of liability, under 200 words |
| Known traps | Where do drafts usually go wrong? | Invented legal references, over-apologising |
| Inside or outside the edge | Can I check this fully? | Inside: I can verify every claim |
| My final check | What do I read before it leaves? | Every date, figure and legal reference, against source |
This is the row people skip, and it is the whole advantage. Write it as if briefing a capable new colleague. Anthropic's "golden rule" is to show your prompt to a colleague with minimal context and see whether they would be confused; if so, the model will be too.
Test it: could you spot a wrong answer within a minute or two? If yes, it is inside your edge for now, and a draft saves you time. If you would have to take the output on trust, it is outside. Either learn the area properly first or use the model to explain and point to sources, not to decide.
Use something like this. Fill in the bracketed parts from your card.
You are a [YOUR_ROLE] assistant working for [YOUR_ORGANISATION] in [YOUR_COUNTRY].
Task: [THE_EXACT_TASK]
Audience and tone: [WHO_READS_IT_AND_HOW_IT_SHOULD_SOUND]
Context a newcomer would not know:
[CONTEXT_ONLY_I_HOLD]
A good result:
[WHAT_GOOD_LOOKS_LIKE]
Known traps to avoid:
[KNOWN_TRAPS]
Source material (use only this for facts, figures and rules):
[PASTE_SOURCE_DOCUMENTS]
Steps:
1. If anything above is missing or ambiguous, ask me before drafting. Do not guess.
2. Draft the result in the format [OUTPUT_FORMAT].
3. List every factual claim, date, figure or rule you relied on, and quote the source line for each. If you cannot find support in the source material, say so and mark the claim "unsupported".
4. Finish with three things I should check myself before using this.
If you are unsure about anything, say "I am not sure" rather than filling the gap.
Fill in the role, task, context, definition of good, traps, source material and format. The "unsupported" step follows Anthropic's hallucination guidance: allow the model to say it does not know, ground answers in quotes from provided documents, and have it retract claims it cannot support (Reduce hallucinations). That page also notes these techniques reduce hallucinations but do not eliminate them.
Read the draft against your "final check" row. Every time you catch a new error, add it to "known traps". The card gets better each time you use it, and that accumulated judgement is yours, not the tool's.
Hypothetical scenario: Priya, a lettings manager in Leeds, spends two hours a day answering repair complaints. She fills in the card above. In step 2 she realises that routine repair replies are inside her edge, but questions about tenancy deposit disputes are not, because she cannot always tell whether the model's reading of the rules is right. So she uses the prompt for the first and, for the second, asks the model only to list what documents she should gather before speaking to a specialist. After a week she has added two traps to her card: invented clause numbers and promising deadlines the firm cannot meet. Her edge is not that she knows the tool. It is that she knew what to put in the card.