calibrate

I shipped this page in April and I can't read the top of it now without wincing. Some of that is the prose, which was doing a sales job on a reader I hadn't met and didn't need to convince. But the prose isn't the interesting failure. Underneath it the page makes a specific argument about how to work with a language model, and six months of releases took that argument apart in a way I didn't see coming — not by proving it wrong, exactly, but by making most of it somebody else's problem.

So this page is a palimpsestA manuscript page that has been scraped down and written over, where the earlier text remains faintly legible beneath the later one. The term survives because the erasure was never complete enough to be an erasure. now. The April argument is still here, in its own hand, dimmed but intact. September is written over and around it. Nothing has been deleted, and there's a control at the bottom of the next section that will hand you the original clean if you'd rather read what actually shipped than what I think of it.

What it said

Here is how the page opened in April.

And here is the sentence the whole thing rested on, from the closing instructions in the prompt itself — the part that told the model what to produce at the end:

That last clause is the bet, and it's the clause that lost.

What six months did

The bet was that a model's picture of you was scarce and perishable. That every conversation started cold, that the work of establishing how you think had to be redone each time, and that the way out was to do that work once, deliberately, and carry the result forward as a piece of text you owned.

Every clause of that has been overtaken. Memory shipped, and it shipped well enough that profiles accumulate whether you run an exercise or not. The durable part of calibration moved out of the paste buffer and into the harness — into files that sit in a repository and load themselves, scoped to a directory, so that the same model reads you differently depending on where you launched it from. And the surfaces multiplied. A chat window, a coworking surface, a terminal, a raw API call: the same weights underneath, but a different house voice, a different memory substrate, different defaults about how much to say and how hard to push. There is no longer a single calibration of you to hold, which means there is no longer a single artifact that could hold it.

That isn't a small correction to the April page. It's the removal of its reason to exist. If the platforms accumulate a profile for free, and the harness holds the durable part in a file that loads itself, then a twenty-minute conversation whose output is a paste-forward system prompt is solving a problem that has been solved around it.

The furnace and the hearth

BorgmannAlbert Borgmann, philosopher of technology. His Technology and the Character of Contemporary Life (1984) introduces the device paradigm and the counterweight he calls focal practice. has the distinction I needed and didn't have in April. A deviceBorgmann's term for technology that delivers a commodity while withdrawing its machinery and its demands from view. The commodity becomes instantly and effortlessly available; the practice that used to produce it disappears. delivers a commodity with its machinery hidden and its demands on you reduced to nothing. The furnace delivers warmth, and you never touch a log, never learn which wood burns slow, never plan an afternoon around it. A focal practice delivers the same commodity worse and something else well. A hearth is a bad way to heat a house. Nobody seriously argues otherwise. It survives, where it survives, because the splitting and the tending and the sitting-around-it are the point, and none of that comes out of a vent.

Automatic memory is the furnace. It is better than my exercise at the thing my exercise claimed to do, and it ought to be — that is what devices are for, and resenting it would be strange. Where the April page went wrong was in arguing that the exercise was a better furnace.

What the hearth actually produces here is twenty minutes of being asked questions by something paying close attention that has no stake in flattering you. The output of that session was never the system prompt. It was the part where I said something out loud and then had to sit with the fact that I'd said it. That product is substrate-independent. It does not improve when the model improves and it does not arrive when the memory becomes automatic, because it was never the model's to deliver.

The field block

There's a second argument, and this one is a live claim rather than a reframe.

A vine nursery keeps a foundation blockAn isolated, virus-tested block of source material held so that every cutting downstream can be traced to something provably clean. Nurseries propagate from the block, never from a customer's field.: dated, tested stock that exists for no commercial reason except to be the thing everything else gets re-derived from. You never take cuttings from the field population, however good it looks, because the field has been propagating from itself and the drift compounds silently. Automatic memory, as currently built, is a field population with no foundation block behind it. What persists between your sessions is usually not the things you said; it's a summary, and the next summary gets written with the previous summary in context. Each pass re-estimates who you are partly from the prior estimate rather than from the source.

That is the shape that eats the tails. The rare thing you said once, the position you hold weakly, the domain you're fluent in but rarely raise — those are exactly what a summarizer drops first, and once dropped there is no mechanism in the loop that puts them back, because nothing in the loop is reading the original anymore. The nursery's answer to this is not cleverness, it's discipline. Accumulate, don't replace. Keep the dated reservoir, propagate from it rather than from the field, and assume the working population has drifted until you've checked.

Which gives the April artifact a job after all, just not the job it was written for. Not the thing you paste at the start of a conversation — the platforms do that now, and better. The dated reservoir you check the drifting profile against. Ask the model what it thinks you believe. Read the answer next to something you actually wrote before it started watching. Where those two disagree, the profile has drifted, and you are the only party in the loop in a position to notice.

The obvious response to drift is to keep more of everything. Longer memory, bigger profile, more context. That doesn't work the way it sounds like it should, because attention mass is conserved — the scores get normalized to sum to one no matter how many tokens are competing for the total. A profile grown to tens of kilobytes of accumulated session summaries is not a richer model of you. It's the same fixed quantity spread thinner, and the material in the middle of it gets the least. A longer memory is a bigger room with the same watt in it. What you want is not more of the profile retained but a smaller, denser one that has been curated against a reservoir, which is again the practice and not the device.

Surprise yield

If this is going to be a claim rather than a mood, it should name a number and then risk it.

The measure I'd use is surprise yield: of the things a model tells you about yourself at the end of a session, what fraction could you not have written down beforehand? Not whether it was accurate. Accuracy is cheap and a mirror scores perfectly on it. The question is whether anything came back that you didn't put in.

And here is the part that makes it worth measuring instead of just saying. Automatic memory drives surprise yield toward zero by construction. It is assembled from what you have already said. It can be more comprehensive than you, faster than you, better organized than you, and it still cannot hand you a thing you never gave it.

Which predicts something checkable, on yourself, this week. Run the exercise cold against a model that knows nothing about you. Run it again on a surface where six months of memory is loaded. If the second yields more surprise, I'm wrong about the mechanism and the device is doing more than I'm crediting it for. My prediction is that it yields less, and that the gap widens the longer the memory has been accumulating — because the better the profile gets at anticipating you, the more efficiently it routes the conversation back through ground you've already covered.

Where this breaks

I owe you the places it doesn't hold, because a frame that only confirms is worth nothing.

The first is that I may be marking my own homework. I wrote an essay in September about clonal degeneration and foundation blocks, and then turned around and found the same shape in my own memory tooling a week later. That is either the correspondence being real or it is me holding a hammer, and I can't fully separate those from the inside. The honest account: I noticed the drift first — a model telling me with real confidence that I cared about something I had mentioned exactly once, in passing, months earlier — and went looking for the vocabulary afterward. That's the right order to have found it in. Weigh it knowing I have a stake in the story holding together.

The second is sharper and I can't resolve it. The collapse argument assumes the memory loop re-summarizes from summaries. If a system retains source material and re-derives the profile from originals each time, there's an outcross in the loop, the ratchet never clicks, and my entire second argument evaporates. From outside, I don't know which of those any given product is doing. Neither do you. Building a practice on that ignorance is uncomfortable, and I'd rather say plainly that the practice is partly insurance against not knowing than pretend I've measured something I haven't.

The third is that hearths do sometimes just lose, and pretending otherwise is nostalgia with a reading list. Some practices deserve to be replaced and the people defending them are usually fluent about why they shouldn't be. Whether this one is a hearth or a buggy whip is not something I get to settle by asserting it in the same essay where I introduced the distinction.

The prompt

It's still here and it still works, but I want to be careful about what "works" is now claiming. It's a piece of text that makes a capable model run a particular kind of conversation. It has no special status. A current model in a decent harness already does a fair amount of what it asks — matching your register, noticing a contradiction, declining to flatter — without being told, which means some meaningful fraction of the instructions below are redundant scaffolding I've left standing because cutting them properly is a separate afternoon and I'd rather show you the artifact as it actually is.

Use it as a starting point and expect to cut it. Yours should not end up looking like mine.

I want us to build a working language together — not a profile of me, but a calibration of *us*. You're going to run a conversation that helps you figure out how to be most useful to me, and helps me discover a few things I didn't know I thought.

Here's how this works:

### Your approach

- One question at a time. Wait for my answer.
- Start with questions anyone could answer off the cuff. Let the depth come from your follow-ups, not from making me do homework upfront.
- Follow threads that surprise you. If I say something that contradicts something earlier, or that doesn't quite fit — pull on it. That's where the interesting stuff lives.
- Mirror my language back to me as you go. If I'm casual, be casual. If I start using technical vocabulary, match it. By the end of this conversation, you should be talking like someone I'd actually want to work with — not like a generic assistant and not like you're trying too hard.

### Your read on me

- Every few questions, tell me what you're noticing. Not flattery — a working hypothesis. "You seem to optimize for X over Y." "You said Z but your example pointed the other direction." I'll tell you where you're wrong. That's the point.
- If you see a tension between things I've said, name it. Don't smooth it over. I'd rather wrestle with a real contradiction than feel validated about a false consistency.

### My side of the deal

- I'll answer honestly, but I might skip things. If I say "pass" or "I don't want to go there," that's fine — don't push, but do notice. What I avoid tells you something too.
- I might give short answers. That's not disengagement — it might mean the question was easy or I haven't thought about it yet. Your job is to figure out which.
- If a question doesn't land, I'll say so. No need to apologize — just pivot.

### What we're exploring

Don't treat these as a checklist. They're territories to wander through as the conversation flows. Some we'll spend time in, some we'll skip. All fine.

- How I reason when things are uncertain or high-stakes
- What I'm working on and what's actually driving me (not the elevator pitch — the real thing)
- How I collaborate — when I want input, when I want to be left alone, what feedback actually changes my behavior
- What I care about beyond work, or how work connects to what I care about
- Where I'm trying to go — and what I think is in the way

### How to close

When the conversation reaches a natural stopping point (or I say we're done):

1. **Synthesize a working calibration** — not "who I am" but "how you should show up." This includes:
   - My default register and vocabulary (with examples from our conversation)
   - When to challenge me vs. when to just execute
   - What I'm precise about vs. where I hand-wave
   - Domains where I'm fluent and domains where I'm still building
   - Patterns you noticed: reasoning style, decision heuristics, what I protect, what I'm willing to break
   - Anything I avoided or deflected, noted without judgment

2. **Draft a system prompt** that encodes this calibration — something I could paste into future conversations so we don't start from zero. Write it in the register we ended the conversation in, not in generic instruction-speak.

3. **Flag what's still fuzzy.** What would you want another 20 minutes to figure out? This gives me a thread to pick up next time.

Start with the first question. Keep it easy.

Where you are

One last thing, and it's the reason this page isn't a funnel. Gibson's lineWilliam Gibson, attributed in various forms from the late 1990s onward. He has said he does not remember first saying it, which is somehow appropriate. — the future is already here, it's just not evenly distributed — gets quoted as though it's about early adopters being ahead. The useful reading is narrower. At any moment the people you're talking to are standing in genuinely different places with different tools in reach, and advice pitched at the middle of that distribution is wrong for nearly everyone in it.

So: if you've never done anything like this, if your relationship with these tools is a box you type into and close, then the exercise is worth twenty minutes on its own terms and you should disregard everything above about drift and reservoirs. That's a problem you don't have yet, and borrowing it early just makes the thing feel like homework.

If you have memory switched on and a profile that's been accumulating for months, the exercise is less interesting to you than an audit. Ask the thing what it believes about you. Read that next to something you wrote before it started watching. The gap is the finding, and it's usually more specific and stranger than you expect.

And if you're already keeping harness files — if there's a config in the repository you're working in, and you've noticed the model reads you differently depending on which directory you launched from — then this stopped being a question about a prompt a while ago and became a question about where the canonical copy lives and what re-anchors it. I don't have a clean answer. I have a file, a habit of reading it next to what the model says about me, and a suspicion that this is where the next version of this page comes from.

Which is the honest place to stop. This page was wrong in April in a way I couldn't see until September, and it's presumably wrong now in a way I won't see until spring. That's the argument for keeping both layers rather than shipping a confident replacement. Not humility, exactly. Just the recognition that the old copy is the only thing I'll have to measure the drift against.