I heard this one from Geoff Woods on Codie Sanchez’s BigDeal podcast. Being honest, it’s a two-hour listen and Woods is borderline insufferable, but the advice underneath is genuinely good: simple enough to explain in a minute, and useful even if you’ve already built big things with AI. So here it is, saved you the two hours. It’s four parts, and they spell CRIT, Context, Role, Interview, Task.
The shape matters as much as the parts. Nearly all of a good prompt is context; the rest is roughly even.
C Context
Give the model everything about what you’re doing, and the moment you think that’s enough, give it more. This is where nearly all the quality comes from, and it’s the bit everyone shortchanges. I dictate this part rather than type it, talking pulls far more out of your head than typing does, and you stop quietly editing yourself down to three tidy sentences. I wrote more about why that works here.
R Role
Tell it who you’d actually want advice from, and get specific, niche even. Not “a business consultant”, but “a strategic consultant who lives in contracts and retainers”. One thing the podcast didn’t say, but I would: don’t reach for “you are a world-class expert”. Controlled studies say it doesn’t make the answer more accurate and can make it worse, all while making it sound more sure of itself. You’re not puffing it up, you’re pointing its thinking. I did a whole write-up on that here.
I Interview
Before it answers, get it to interview you, one question at a time so each answer can shape the next. The podcast said define three to five questions; I just leave it open and let it run until it’s got what it needs. It’s the step that turns a decent prompt into a great one, because it fills the gaps you didn’t know were there. There’s a paste-ready version of this here.
T Task
Then, and only then, the actual job. What does the finished thing need to be? Keep it to the outcome, you’ve already done the heavy lifting above.
The shape, assembled
Context: [everything about what you're doing — dictate this, don't type it] Role: act as [the specific person you'd want advice from, not "an expert"] Interview: before you answer, interview me one question at a time until you have what you need Task: [the finished thing you want]
A real example: the retainer
Last week I used exactly this to draft a retainer for the company I’m leaving, so they can still get my help after I go.
- Context: I dictated for ages, everything I’d been doing, where I thought they’d get real value, how I wanted it to work.
- Role: a strategic business consultant who specialises in contract and retainer work, not a generic business guru.
- Interview: I left it open. It asked about twelve questions, from the shape of the business to what I wanted out of the arrangement, all sharp.
- Task: write the retainer.
The draft that came out was miles better than anything I’d have typed cold.
The bit I wouldn’t skip: peer review
After the draft, I had it spin up a few subagents to peer-review it and hunt for the gaps, three passes. They caught things I’d have happily shipped: no minimum spend on the ad-hoc extra time, and me quietly agreeing to cover my own travel. That’s the difference between a prompt and a process. If you’ve not set subagents up, start here.
That’s it
And if you’re reading this thinking you’re past frameworks, honestly, so was I. I don’t think in frameworks when I’m building, I don’t need to. But every time I go back to having it interview me before it writes anything, I’m shocked by how good the questions are.
Context, Role, Interview, Task, and a peer review to finish. There’s a video coming on this, but you’ve got the whole thing here first.
If you try it, reply to the email or drop me a message and tell me what you built, I read everything.