Communicating with AI
Saying what you actually want, to something that takes you literally.
The behaviour, in one line: I get strong results from AI by giving clear context and examples, iterating, and verifying the output.
A very fast, very literal contractor builds exactly what is on the drawing — including the parts you did not mean to draw and the parts you left blank. You are not making it understand you. You are removing the places where it has to guess.
This is a communication skill in technical clothes. A model cannot see what you meant, only what you wrote, and it will fill any gap you leave with something plausible. What comes back is bounded by how clearly you said what you wanted, what it is for, and what a good answer would look like. The side effect is that practising it makes you clearer with people, who have been quietly filling your gaps for years.
| What it is | What it isn’t |
|---|---|
| Stating what good looks like | Assuming it knows what good looks like |
| One worked example | Three paragraphs of description |
| Correcting one thing at a time | Rewriting the prompt from scratch |
| Checking the output | Trusting how it reads |
What it looks like. The first answer is nearly right and wrong in one specific way. Rather than rewriting the whole prompt, you say which part is wrong and what it should be instead. The second answer is usable.
- Any task you would otherwise have delegated. The brief is the work. A vague brief produced vague results long before models existed.
- Work where being wrong is expensive. The output is equally confident either way, so the checking has to come from you and it has to be routine.
- Getting past the demo. Almost anyone can get something impressive once. Getting something reliable repeatedly is the part that is actually a skill.
- Say what it is for and who it is for. Most disappointing output is a reasonable answer to a question you did not ask.
- Show one example of the output you want. An example does more work than three paragraphs describing it.
- Steer across turns rather than rewriting from scratch. Correcting one thing at a time is faster, and it tells you which thing mattered.
- Check the result every time. It reads the same whether or not it is right, and that fluency is what you are least able to resist.
- Break it into parts
Four parts: supplying context and constraints, showing an example, steering across turns, and checking the result.
- The part that carries it
The worked example. One example of the output you want does more than three paragraphs describing it.
- The order to learn it in
- Start by always stating the goal, the audience, and what good would look like.
- Then add a worked example of the output you are actually after.
- Then practise steering across turns — correct and narrow rather than starting again — and check the result every time.
- Put something at stake
Take one prompt you use weekly, rewrite it with context and an example, and show a colleague both outputs side by side.
- Compress it
Say what it is for. Show what good looks like. Steer across turns. Check the result.
- The smallest useful dose
One recurring prompt upgraded and saved each week.
- Anchor it to something you know
Briefing a very fast, very literal contractor who has never seen your building. They build exactly what you drew, including the parts you drew by accident.
“The hottest new programming language is English.”
— Andrej Karpathy
The interactive version is in the course. The diagram, the simulation and the four-question check for Communicating with AI run in the free course, where your progress across all 24 skills is kept.
Open Communicating with AI in the course →
Not sure where you stand? Take the twelve-minute self-assessment, or see the 30-day course that turns these skills into a roadmap.