Producing a thousand words costs nothing now. Reading a 1k words still costs 4-40 minutes and cognitive power.
A contractor sent me a proposal last week. The covering message said he’d kept it short, just like I asked. It was in ant-sized font to keep it under 1 page. Typical high school trick.
The content was good. Three criteria for scoring creator content, all sensible, he’d clearly thought about the problem. But it took me re-reading it three times in two different sessions to understand the core mechanic of the proposal.
Humans don’t structure thoughts like that. Humans usually think “what do I want to tell this person” rather than “how do I cram all the context from this 10k char back and forth chat into 1k chars”.
The standard response here is to ban the tells. No em-dashes, all lowercase, no triads. I get the impulse, I’ve seen teams write the rules. But telling AI to use lowercase and banning em-dashes doesn’t help this aspect too much. Strip every tell out of that document in one pass and it stays exactly as hard to read — even harder perhaps, as we’ve all learnt to skim AI slop faster than reading carefuly prose (partially because it repeats itself so much).
An AI model optimizes for coverage: every relevant consideration from the 300k token chat, condensed into its decisions, arranged flat. A person writing to a busy person does close to the opposite. They lead with the thing that changes your decision and drop the rest, because they have a model of you. That is the work today.
It’s also the easiest part to miss, because when the prompter reads the output, it looks finished. In your head you might have a shortlist of what you want to see in the output. Budget? Check. Timeline. Check. Criteria list? Check. Details about each criterion? Check. Criteria looking correct? Check. The problem is that the checklist is a bool on each item. So maybe you could some noise on the items that didn’t make your shortlist. But humans order things hierarchically. So the real question is not “is this covered” but “how is this relevant in the idea maze” and “does the reader have to know”.
For that contractor turns out it his 2-page writeup could have been 2 short bullet points. The two pages went over each milestone (which were standard milestones like “planning, implementation, verification” stuff that is like mentioning you will get a spoon and a fork with your potato salad and a spoon for your tomato soup (although coperto is a thing in Italy lol).
Something like this:
Two weeks. $2k management fee, $5k budget for creators. Max total $8k. >20 creators, I define quality bar and you get veto on each. Approximately 40 posts, mostly video, clippable for later.
Objective: testing whether the programme could scale to $50k.
Thirty seconds to read. I know what he thinks. I can say yes or no.
When the cost of producing something goes to zero, the price moves to whatever didn’t move with it. We are paying people for exactly two things:
human intelligence
human connections
That’s the entire list. Everything else on the invoice is now a commodity. Obviously this is self-referential as the part of “human” includes the stuff that is not commoditized and stuff is commoditized because it’s not human. However - that’s the first step for asking “what is it that cannot or should not be automated in the knowledge economy”.
On (1 - human intelligence): what I can’t get from a model is context I never had to state. When I learn something it arrives attached to who told me, how they said it, what shoes they had on, and what happened the last time we tried it. None of that is in the prompt, it can’t be no matter how many creepy RayBan glasses you attach to your body. Some call it taste.
On (2 - human connections): a relationship is someone accepting a liability on your behalf. When a person vouches for you they put something at risk that can actually be taken from them. A model can write the intro email. It can’t be the one who loses face when the intro goes badly. That reputation staking system is uniquely human for now, bot identities don’t have reputation in the social sense.
So, if you’re the one sending: length now a liability. Every paragraph you didn’t cut is a claim that your time is worth more than the reader’s, and you’re making that claim ten times a week, to people who are counting.
If you’re the one receiving: stop testing for AI. Wrong test and a whackamole. It’s not that AI is bad. It’s that it flips the asymmetry in who does the cognitive work. At Peanut, we use AI a lot but at the same time we are extremely cognizant of the fact that it’s easy to waste team member time with extremely verbose writeups. In other words: it’s not about stylistic choices, or someone being angry about em-dashes. It’s mostly about who spends their very limited neuron tokens.