📩 The Five Questions Before You Hand Anything to AI / The Best AI Policy Never Mentions AI
20 August 2026
buildwitholi
[AGENDA] In today's newsletter
- [UPDATE] Where I am this week
- [FRAMEWORK] The five questions before you hand a job to AI
- [WRITING] The best AI policy never mentions AI
- [TECHNIQUE] Stop letting one expensive brain do the legwork
- [THE COURSE] The group version, in October
Hey,
[UPDATE] Where I am
I'm mid-move at the moment, which has landed me at my dad's for the best part of a week. Long dinners and chats in the garden, what's not to love.
[FRAMEWORK] The five questions before you hand a job to AI
The AI Daily Brief put out a scoring method for how to decide which use cases are good to hand over to AI, and which are not.
Take a job you do regularly and score it zero, one or two on five questions:
| Two if | Zero if | |
| Is it worth it | it's weekly and it drags | twice a year, ten minutes |
| Can you teach it | ten minutes on a screen share would do it | it only lives in your head |
| Can you check it | you can tell at a glance | you have to redo it to know |
| If it's wrong and nobody notices | it's an internal draft | it reaches a client, a regulator or your accounts |
| Are you the reason it's good | anyone competent would match it | it's good because it's you |
Then, add it all up /10:
| Total | Verdict | What to do |
| 8 to 10 | Deputise | Hand it over. Spot check the first few and then leave it alone. |
| 4 to 7 | Duet | Split it. You keep the thinking, AI does a defined piece of it. |
| 0 to 3 | Defend | Keep it. Either the mistakes are expensive or the person on the other end is expecting you. |
The one that catches people out is the third question. Plenty of jobs are quick for AI to do and slow for you to check, and those feel like wins right up until you're proofreading everything it gives you.
I thought this was a really tidy way of thinking about it, so I hope you do too. I've written it up in full, and there's a video coming on it soon, so you're getting this one early.
[WRITING] The best AI policy never mentions AI
We've all sat in a meeting where someone is presenting something they're genuinely proud of, and then a hard question lands and it turns out AI wrote that bit. Proud to disowning it in one question.
Also off the back of the AI Daily Brief, a company published its AI writing policy and the interesting part is what isn't in it. Three rules, and not one of them mentions AI:
- You stand behind every line. If there's a sentence in there you couldn't defend when someone stops you on it, it doesn't go out.
- Writing is thinking. The argument that felt solid in your head falls apart on the page, and the falling apart is the useful bit. Draft with AI, argue with it, get it to attack your reasoning. Asking it for the finished thing and sending the finished thing is the part that costs you.
- It should take longer to write than to read. If five minutes of your time creates an hour of everyone else's, you've moved work off your desk and onto theirs.
They're the things writing used to enforce on its own, back when producing ten pages actually cost you something. That's gone, so somebody had to write them down.
[TECHNIQUE] Stop letting one expensive brain do the legwork
By default your AI does every part of a job itself, the thinking and the grunt work, all in the same conversation. That's the senior partner renaming files, and it's why long chats get slow, expensive and forgetful.
It can hand pieces of the work to helpers instead. Worth knowing that a subagent runs on whatever model you are on unless you say otherwise, so naming a smaller one is where the saving actually comes from.
One for research, when you want the reading done without it filling up your chat:
[INSERT CONTEXT] Send subagents out to research this on Sonnet or Haiku, and bring back [INSERT GOAL]
And one for building, when the job has several parts to it:
[INSERT CONTEXT + GOAL] Plan this yourself first, then hand each part of the work to subagents on Sonnet or Haiku, and review what they give you before it lands.
This one has been around for a while, but it's one that I keep coming back to, so here's the full guide along with how to make it permanent.
[THE COURSE] The group version runs in October
The other bit of news I'm genuinely pleased about, about half a dozen people have signed up to test the agent ecosystem course with me one to one next week, which is amazing. That came from the last email, so thank you. I look forward to letting you know how they found it.
This is being done to practice and refine the content before I run it as a course for a small group across four weeks in October.
If you're interested, put your name on the list below!
And if you've got any questions about it, just reply to this email.
Chat soon,
Oli