BACK TO RESOURCES
Understanding AI

Three rules for writing with AI

Not one of them mentions AI. They are the things writing used to enforce on its own.

beforenoweffortpages produced

Writing used to police itself. Ten pages took a long time, so length was decent evidence that somebody had thought about it. That link is gone, and the rules below are what is left once you cannot read effort off the page count any more.

What you'll walk away with

  • The three rules, and the reason each one holds up
  • Why not one of them mentions AI, and what changed underneath to make them necessary
  • The honest note on what a policy like this will not fix

About five minutes to read.

Why each one holds

1

You stand behind every line

This is the one that does the work. The failure everybody recognises is not that somebody used AI, it is watching them present something they are genuinely proud of, and then the moment a hard question lands, it turns out the AI wrote that bit. Proud to disowning it in one question.

The rule closes that exit. It does not ask you to write less with AI, it asks you to read what you send. If there is a sentence in there you could not defend if someone stopped you on it, it does not go out. That is the whole test, and it is the same test that applied before any of this existed.

The practical version. Before you send anything, pick the two lines a sceptical reader would push on first and check you can actually answer for them.

2

Writing is thinking

Writing something down is how you find out whether you understand it. The argument that felt solid in your head falls apart on the page, and the falling apart is the useful part. It is why the document is worth producing at all, and it is the bit AI can do for you.

So if you hand the whole job over, you get the document and you skip the understanding. You will not notice until someone asks you a question that sits half an inch outside what the AI wrote, and you have nothing behind it.

This is not an argument against drafting with AI. Draft with it, argue with it, get it to attack your reasoning and tell you what you have missed. All of that is thinking. Asking it for the finished thing and sending the finished thing is the part that costs you.

3

It should take longer to write than to read

This is the one that stops the flood, and it is the easiest to check. Time spent writing on one side, time spent reading multiplied by the number of readers on the other. If the second number is bigger, you have moved work off your desk and onto everyone else's.

Ten pages, ten readers

You, writing it5 minutes
Them, reading it2 hours

Which gives you the most useful line in the whole policy. If the entire thing came out of a one line prompt, just send the prompt. Your reader can generate it themselves in the shape they actually want, and nobody has to pretend to have read ten pages.

Notice what is missing

Not one of the three rules tells you when you are allowed to use AI. There is no approved tools list, no percentage, no declaration form. They are basic principles about owning your work, and they were true long before any of this.

They only need writing down now because of what changed underneath. Producing a long document used to be expensive, so length carried information. Nobody wrote ten pages they had not thought about, because ten pages of nothing was more work than saying the thing properly. The cost did the policing, quietly, and nobody had to make a rule out of it.

That cost is gone. Effort and length came apart, and once they did, every judgement that used to ride on length stopped working. So the thing effort used to say on its own now has to be said out loud. That is all a good AI writing policy is, and it is why the best one going does not mention AI.

What stopped working

it is long, so it is consideredbecomescan they answer for line four

What this will not do

It will not catch anyone. There is no test in here for whether something was written by AI, and that is on purpose, because you cannot reliably run that test and a policy built on catching people turns into a game of hiding it better. Every rule here is enforced by a person asking a question in a room, which means it only works if people actually ask.

It also will not survive being posted and never mentioned again. The one behaviour that makes it real is a manager pushing on a line and expecting an answer. Do that twice and the policy holds. Skip it and you have a document nobody has read, which is the exact thing it is meant to prevent.

Where this came from

I heard about this on the AI Daily Brief, which is worth your time if you want one podcast on this stuff. The policy is Clay's. It was written by an engineer there, Sophie Alpert, for the engineering team, other departments asked to use it, and their co founder Varun Anand made it company wide.

The wording above is mine, and I have folded their fourth rule about length into the third, because it is the same idea and it lands better as one. Crediting it properly seemed like the least I could do on a page whose first rule is that you own what you send.

Next