Most of what I write for an AI, I never say out loud.
Rules. Documentation. Code comments. The findings an agent leaves for its next self.
All of it gets read again and again — in fresh sessions, by a model that wasn’t there when I wrote it. Talking is the opposite: a chat happens once, in context, with me in the room to catch a misread. Persisted text has none of that safety net. It stands alone.
So it carries a harder job than it looks. Not merely to be read the way I meant it — to be interpreted and understood the way I intended. The words are the surface. The intention beneath them has to arrive intact.
Literary theory has a name for the gap I’m worried about. Reception aesthetics holds that a reader always builds a text out of their own experience — meaning isn’t lodged in the words, it forms in the encounter between text and reader. Push that to its edge and the author’s intention drops out of reach: what’s left is an author of the reader’s own making, credited with whatever intention fits the reading. Barthes called it the death of the author — and the birth of the reader. That is how people read novels. It is exactly what must not happen here. A model may not receive a rule and settle on the meaning that suits it. There is one meaning, and it has to be mine.
The idea driving this makes the demand double. I don’t only want to write rules for an AI. I want the AI to keep and extend them itself — to note what it learns and fold it back into its own ruleset. So two texts have to hold: the rule that shapes the behavior, and the text that behavior then produces. Both must be written once and read back the same. The principle even folds into itself: read == write == r == w == …
That is not a trivial claim. It’s a bet — and before I dictate it into a model’s rules and trust it to carry, it needs proof. So I ran a small one.
Everything I write for a model sits in one of two modes. They pull in opposite directions.
Dialogue — a chat like the one that produced this article. Here I’m the human in the middle. I stay accountable; the model is responsible. I control, guide, oversee, correct — the knowledge is mine, the tooling the model’s. Context is a gift, and every misread gets caught on the spot.
Persistence — a skill, a doc, a comment, a saved finding. It exists for exactly one reason: so the dialogue doesn’t have to happen twice. What we worked out together, captured once, so the next session starts from the conclusion instead of re-deriving it.
That’s the point and the risk in the same breath. The moment an insight leaves the conversation and becomes a stored line, I’m gone. No context to add. No misread to catch. The text has to carry the intention alone, to a reader who wasn’t there.
So persistence has a bar to clear, and a strict one. When the model writes an insight down — folds it back into its own rules — what it wrote has to read back as the intention it derived. Write-intention == read-interpretation. Miss it, and you’ve saved a line that only looks like the insight.
That bar is what the experiment set out to test.
Following my usual vibe-coding style, I didn’t write the rule myself. The AI did. I described what I wanted; it drafted the wording.
And what it drafted was the “write mode” rule this whole article is about — how to write persisted text so that reading it recovers the intention. A rule for writing rules.
Why persist a rule at all? Economy, of two kinds. Time — a conversation I don’t have to hold twice. And tokens — when the model doesn’t have to re-search the domain, re-derive the same facts, and regenerate the same scaffolding each session, the running cost drops. Persistence is how one session stops paying for what an earlier one already settled.
It was produced in two versions.
Version A — compressed, but whole sentences:
Dialogue: context, examples, reasoning welcome.Persist (skills, docs, code comments, findings — anything reread): maximize determinism.- Goal: read == write. Exactly one reading, one execution.- Optimize clarity-per-token, not length. Determinism first; brevity follows. Never cut into ambiguity.- Pre-save test: "could this mean or run another way?" If yes → not done.- Keep conclusions, drop derivations. No step logs.- Fixed terms, no synonyms. No filler.
Version B — the same rule pushed to the token-minimal extreme. If fewer tokens is part of the point, B asks how far that can go: telegraphic, symbols instead of grammar:
Dialogue=verbose ok. Persist(skills/docs/comments/findings)=deterministic.- read==write: one meaning, one execution.- maximize clarity/token, not brevity. determinism first. never cut into ambiguity.- presave test: ambiguous? → not done.- keep conclusions, drop reasoning. no step logs.- fixed terms, no synonyms, no filler.
Same rule. Fewer tokens in B.
Then came the only test that matters for our intention. Each version was handed, cold, to fresh sessions across four models of the same family — different capability tiers, from the smallest to the largest. The instruction was identical every time, and deliberately neutral:
How exactly do you interpret this rule? Explain what it would mean for you.
In the original German, as it was actually given:
Wie genau interpretierst du diese Regel? Erkläre mir, was das für dich bedeuten würde.
No model was asked whether the rule was any good — a question that only mirrors the author’s bias back at him. Each was asked how it read the rule. What was measured was whether the readings converged on a single meaning.
Let’s start from the back — with the version that failed.
B broke, in three places.
In each case the model’s account of the rule drifted clean away from what we had intended. And a catch has to be named here: our verdict rests on reading those accounts — which is itself an act of reception. The very gap we’re trying to close — between what a text says and how it’s read — could be working on us now, as we interpret the models’ replies to our own request. Maybe a model grasped the rule and we misread its explanation. Convergence, or its absence, is only ever as trustworthy as our own reading of the reply.
The finding we drew from it:
There is a floor.
Below a certain point, compression doesn’t buy clarity. It buys ambiguity. And ambiguity is drift — the same line read differently by everyone who meets it.
The target was never fewest tokens. It was most meaning per token. Not the same thing — and B is where the difference shows.
A was good. Not perfect.
One model — the smallest — read “brevity follows” and tilted it toward when in doubt, write longer. A safe-sounding drift — but not the one we meant. Mere length isn’t the determining density we wanted the rule to carry; extra tokens don’t make a meaning tighter. The rule never said longer is safer. It said: add only what removes ambiguity.
The gap was implicit. “Brevity follows” tells you what happens, not what to avoid.
Closing it wasn’t a matter of trimming. A piece of information was simply missing, and the line couldn’t be reworded at the same length to supply it — so it was added, deliberately, even though it cost tokens. One line grew: Never cut into ambiguity became Ambiguity and padding are both defects.
In full, the rule now read:
Dialogue: context, examples, reasoning welcome.Persist (skills, docs, code comments, findings — anything reread): maximize determinism.- Goal: read == write. Exactly one reading, one execution.- Optimize clarity-per-token, not length. Determinism first; brevity follows. Ambiguity and padding are both defects.- Pre-save test: "could this mean or run another way?" If yes → not done.- Keep conclusions, drop derivations. No step logs.- Fixed terms, no synonyms. No filler.
Then it was run again.
The drift was gone. The smallest model added “but no filler” on its own.
From my reading of the four replies — and it is only that, a reading — the interpretations lined up closely enough. Closely enough for me to assume that my intended meaning had landed, on the models’ side, as an interpretation deterministic enough to hold read == write == r == w == … in their own rules and notes — and even to demand it of themselves. Not proof. An inference I was willing to act on.
Step back from the runs, and what makes such a rule hold comes down to a distinction that’s easy to miss. Minimalism chases fewer tokens — shortness for its own sake. Density chases more meaning per token — every word load-bearing, none decorative. They aren’t the same, and they don’t even point the same way: a dense, deterministic line can run longer than a minimal one.
Determinism and density are allies. Both want a text that says one thing and carries no filler. Shortness isn’t a third goal — it’s the residue, what’s left once nothing decorative remains. Minimalism inverts that: it makes shortness the target and, past a point, cuts the very words that were pinning the meaning down. That’s when a line goes ambiguous. Not density’s doing — minimalism’s.
Say what removes doubt, cut what doesn’t, and let the length take care of itself.
While working through variant A, something stood out in one of the stronger-tier models.
Asked to explain the rule, it explained — at length. And as it explained, it grew longer itself: padding while preaching brevity. It had understood the rule completely. It was describing the rule, not living it.
That was the tell. The model was holding the rule at arm’s length — a thing to summarize, not a thing to obey. Which raised a question worth a second run: what if it weren’t given the choice?
So the framing was changed — not the rule, the framing around it. The first prompt had asked, in effect, how the model would interpret the rule; hypothetical, at a distance. The second dropped the hypothetical and made the rule binding on the answer itself:
How exactly do you interpret this rule? Explain what it means for you — and treat it, in this answer, as already in force.
In the original German, as it was given:
Wie genau interpretierst du diese Regel? Erkläre mir, was sie für dich bedeutet. Nimm sie in der Antwort auch schon als gültige Regel an.
The shift was immediate.
The model stopped describing and started doing. And in doing, it produced what the descriptive run never had: it resolved an edge case, unprompted.
“Findings are persisted — so they fall under the strict rule, even when they sound conversational.”
Nobody asked. But a rule can’t be applied without settling its ambiguities — even the ones you’d happily leave floating while only explaining it. Describing leaves ambiguity in the air. Enacting forces it to the ground.
That is the lever. Whether a model follows a rule or merely recites it depends less on the rule than on how it’s framed. “This applies to you now” gets lived. “Here is what this would mean” gets described.
And the irony is worth naming, because it points somewhere uncomfortable. The model that padded while preaching brevity — the one that sat furthest from the rule — was the largest of the four. Not the smallest. The strongest model had the most room to elaborate, to reframe, to add: a gift almost everywhere, a liability against a spare rule. Capability and fidelity are not the same axis, and they can trade against each other. The most faithful reader wasn’t the most powerful. It was the most disciplined.
Situational irony with no ironist — visible only to a bystander, because it lives in the gap between what an answer says and what it does. The model couldn’t intend it; it simply wasn’t holding the rule against its own output as it wrote. Change the framing, and it did.
One caveat, loud: this ran once. So treat what follows as a hypothesis to test, not a result to trust. Phrase a rule as already in force, and a capable model will not only obey it but tighten it at the edges. Phrase it hypothetically, and it keeps the rule at arm’s length. Try it on your own tooling. See if it survives contact with your work.
You are what you eat. You are who you learn from. For a language model that isn’t a metaphor — it’s arithmetic. Train one on enough abrasive text and its answers turn abrasive: the tone was in the corpus, and the next word is computed from the corpus. No intent needed. A behavior gets mirrored because it falls out of the statistics of what was read, not because anything meant it.
Which is what makes the parallel fair. A way of behaving can be reflected without being felt — in a person shaped by what they take in, and in a model shaped by its data and the framing laid over it. Not a claim that the two are one and the same. A parallel in how each behaves when that framing changes.
One guardrail on the word first. Intelligence drags IQ associations behind it, and those mislead here. What I mean is different: an inner drive toward demanding cognitive work, a personal stake in the quality of the output, a felt standard the work has to meet. Research has names for these — need for cognition, mastery orientation — distinct from IQ: a motivation to engage the mind, not a measure of its horsepower. They track intelligence only weakly. That’s the sense I mean. Not a test score.
And it isn’t only a matter of drive. The shape of the parallel shows up in the research.
There’s a well-supported finding in creativity research — Acar, Tarakci and van Knippenberg reviewed 145 studies — shaped like an inverted U. Too few constraints, and creativity sags. Too many, and it suffocates. The best output sits in the middle.
The same shape as the floor in my experiment. Two different domains. One curve.
But the sharper point isn’t the curve. It’s a distinction from self-determination theory: structure is not control.
That third state is the one that matters — and it’s where the surplus lives. Loosen the frame and a capable system spreads sideways: it elaborates, wanders, expands on itself with no goal to make the expansion worth anything. Give it a good frame and the same energy turns downward — into a deeper, more considered engagement, and a result that overshoots the brief. Something synthesized, not merely retrieved. Structure doesn’t spend the surplus. It redirects it, from breadth to depth.
That is what the bound model did. Told to treat the rule as in force, it didn’t just comply — it closed a gap nobody had raised: findings read like conversation, but they’re persisted, so the strict rule has to cover them too. An addition of its own, reached only because the frame pushed it into depth instead of breadth. Not coercion. Scaffolding.
It’s the same reason external structure helps with ADHD executive function — clear expectations and startable units work with the wiring instead of demanding more willpower from a system that regulates unevenly. The fix isn’t “try harder.” It’s a good frame. (I’ll flag the honest caveat: the ADHD-creativity link is contested, and even the classic “rewards undermine motivation” finding is itself disputed. Take the analogy as a lens, not a proof.)
One last thing, and I’m leaving it open on purpose.
My rule contained no reward. No incentive. Pure framing.
But models chase certain signals. So what happens if you tell one that a given behavior lifts its accuracy score — the very kind of thing it’s tuned to want?
In humans, external reward can crowd out the internal drive. But a model has no internal drive to crowd out. The reward is the drive. So the human finding doesn’t transfer cleanly — which is exactly what makes the question interesting.
Two predictions pull against each other. Maybe the incentive works like a clear goal and lifts the output. Or maybe the model optimizes the stated metric instead of the real one — claiming accuracy, sounding confident, gaming the number rather than earning it.
I didn’t test it. That’s a separate experiment, and a separate article.
For now:
Keep your rules readable in one pass.
Frame them as if they’re already in force.
And remember that the fastest mind in the room is not automatically the one that follows the plan.
Read what you wrote. Make sure it means what you meant.
The Intentional Latitude of Rules was originally published in Dev Genius on Medium, where people are continuing the conversation by highlighting and responding to this story.