Working with AI, my field guide

How to work well with AI.
The unknowns method: getting the most out of a model before it starts working.

The gap between what is in your head and what you actually wrote is where things go wrong. Today's models are smart enough; the real skill is noticing what you never said. That skill can be learned, and this page is all of it.

The Four Quadrants

01

Everything about your task falls into one of four quadrants. Picture a kitchen renovation:

Known knowns

What you said out loud

Things you want and actually said. "A white kitchen with gray countertops": it is written in the request, the model knows it.

Known unknowns

Open questions you are aware of

Decisions you have not made yet, but at least you know they are waiting. "Built-in oven or freestanding?"

Unknown knowns

"I'll know it when I see it"

So obvious to you that you never said it. But a drawer that sticks? You will feel it in a second.

Unknown unknowns

The most dangerous quadrant

Things you do not even know exist, so you cannot ask about them. They surface late, and late is expensive.

The best users are not the ones who know everything. They are the ones with few unknowns who assume there are always more, and hunt them down systematically before they blow up.

Stage A, before the work

02

Five techniques that save hours of rework. This is where most of the magic happens.

1Blind Spot Pass

Entering a domain you know nothing about? Instead of guessing, say the truth: "I don't understand this. Tell me what I don't know." Say who you are and what you already know, and use the exact phrases "blindspot pass" and "unknown unknowns"; the model understands them well.

"I'm working on something new and I know nothing about how this domain works. Give me a blindspot pass: help me identify my unknown unknowns, and help me phrase a better request."

2See it before you build it

"Beautiful and clean" cannot be explained in words, but it is easy to recognize by eye. Ask for a sketch before the real thing, ideally several directions at once. Golden tip: ask to see plans and sketches as an HTML page instead of a long wall of text.

"Build me one HTML page with four completely different design directions, and I'll tell you what I like about each."

3Let it interview you

You do not know what you forgot to say, so flip the direction and let the model ask. Around three rounds for a normal task. If you adopt only one technique from this page, make it this one.

"Before you start, interview me: one question at a time, about anything unclear. Start with the questions whose answers would change the direction most."

4Bring a reference

When words fail, show. A screenshot of a site you love, a document structured the way you like, a past result that worked. For the technical: the best reference is source code. Point at it and say "make one like this".

"Attached is an example of what I want. Look at it and build mine in the same style."

5A work plan, but a smart one

Without guidance, a plan mixes the decisions you care about with technical details that mean nothing to you. Tell the model how to sort it: what you might change goes on top, the bolts go at the bottom.

"Prepare a work plan as an HTML page. Open with the decisions I'm most likely to change: user experience, design, behavior. Put the internal technical details at the end; on those, I trust you."

Stage B, during the work

03

Done planning? Open a brand-new conversation and attach only the approved plan and sketch. Then: one technique that prevents surprises.

Implementation notes

Surprises always come up mid-work, and the model decides on its own what to do with them. A work journal turns every such decision into a record waiting for you: zero surprises, and a ready-made lesson for the next project.

"Keep a journal file named implementation-notes.md. If you hit something unexpected that forces you to deviate from the plan, always pick the more cautious option, log it under 'Deviations', and keep working."

Cost tip: do the thinking with the smart, expensive model. Once the plan is locked, hand execution to a cheaper, faster one. Pay for the expensive brain only when you actually need it.

Stage C, after the work

04

1A summary-and-pitch document

Whoever approves your work starts exactly where you started: with unknowns of their own. Give them the result first, then the answers.

"Package the sketch, the plan and the notes into one document I can send for approval. Open with a demo of the result, then answer in advance the questions people will probably ask."

2A comprehension quiz

After a long run, the model sometimes did more than you registered. The rule: do not approve until you pass a short quiz, with a perfect score.

"Prepare a summary report as an HTML page: what changed, why, and how it works now. At the end, add a short quiz about the changes, one I have to pass."

"AI is making us dumber"? That is entirely up to you. "You did something cool? Now teach me how", and the model turns from a quiet worker into a private tutor.

Cheat Sheet

05

Eight sentences to copy as they are:

You don't know the domain"Give me a blindspot pass and tell me what I don't know I don't know."
You want to see before building"Build one HTML page with several completely different directions, and I'll react."
Not sure you said everything"Interview me, one question at a time, starting with what changes the outcome most."
Hard to explain in words"Here's an example. Look at it and do mine in the same style."
Before building starts"A work plan as an HTML page: my decisions on top, technical details at the end."
While the model works"Keep implementation notes. Any deviation: pick the cautious option, log it, continue."
You need someone's approval"Package everything into one document. Open with a demo, pre-answer the questions."
When the work is done"A summary report ending with a short quiz on me. No approval until I pass."

Based on a working method shared by an Anthropic engineer, and on Eliron Giny's Hebrew explainer. The gap between what is in your head and what you said is the whole story. What you learn becomes the map for next time.