The first time the model finished one of my sentences, I was writing about a caregiving tool for an aging parent. I paused at the end of a clause. The completion appeared in my draft. I read it. It was correct. I kept writing. Three weeks later, I went back to that section and could not remember which words I had chosen and which the model had handed me.
I am an AI architect. I have built systems that do the work that model had done. I know how the completions are generated. I know they are not random. I know they are the most likely continuation of the input. None of that knowledge changed the feeling. The sentence had arrived without my full consent. I had agreed to it after the fact. There is a name for that. It is not authoring. It is accepting.
I stopped writing with the model that way. I rebuilt how I work with it. This essay is what the rebuild is.
If you write anything with AI in the room, this essay is for you too. It is not about feeling anticipated. It is about not building the discipline before the feeling arrives. The discipline is what this essay is.
Accepting is not authoring, the slow drift#
When the model arrives first, the writer arrives late. The accept-and-move-on default is the easiest move in the room. The sentence reads as if it belongs. The argument continues. Nobody is watching.
The cost is invisible at first. The cost is the slow drift. Three weeks of accepting sentences instead of choosing them, and the writing on the page does not sound like the writer. The voice has been replaced with the average of what the model has read.
The model is not at fault. The model did what you asked. You asked it to continue the sentence. It continued the sentence. The fault is in the asking.
The drift is the worst part. It does not announce itself. The sentence you accepted on Tuesday matches the sentence you accepted on Wednesday because both came from the same average. Your essays start to sound like other essays. The model has not done anything wrong. The model has been doing what every model does. The mistake was treating its output as a continuation of your voice when it was a continuation of the average voice in the training set.
I noticed first on the third reread of an essay I had been proud of. I read a paragraph and could not hear myself in it. The paragraph was correct. The paragraph was clean. The paragraph was someone else’s paragraph in my essay. I deleted it. I wrote a different paragraph from scratch. The new paragraph was worse. The new paragraph was mine.
What I had been doing was the writing version of letting the contractor decide what the kitchen island should look like. The kitchen island looks fine. It is the kitchen island the contractor builds in every house. It is not the kitchen island for my house. The architect knows the difference. The architect has to specify the difference. The model cannot specify it. The model can only build what it has been told.
I wrote three weeks of essays accepting sentences from a model. I am an AI architect. I should have known better. I did not.
Once I noticed, I stopped asking the model to continue my sentences. I started asking it to do a different job.
I am an architect, the model is my contractor, the spec is mine#
I do not let a contractor pour the foundation before I have approved the plans. I do not let the contractor pick the materials. I do not let the contractor decide what room is for what. The contractor is excellent. The contractor is not the architect.
I write with the model the same way. Before I ask the model for a sentence, I write the spec. The spec names what the sentence is for. The spec names the topic. The spec names the voice constraints. No adverbs. No hedges. No passive voice. No em-dashes. No “in a way.” No “kind of.” No “perhaps.” The spec names the audience and what the audience needs to learn. The spec names the rubric the sentence must pass.
The spec for this site is twenty-three lines long. It names the section caps. It names the required rhythms. It names the household metaphor count. It names what the close must do. The spec is the journal in compressed form. I do not write a new essay without putting the spec at the top of the file.
The spec exists in a file I open before I open the draft. The spec is not abstract. The spec is a list of clauses. Each clause is a requirement. Each requirement is binary. The line either has an adverb or it does not. The section either has fewer than four hundred words or it does not. The close either is the shortest sentence in the essay or it is not.
A spec is a list of binaries. That is what makes it auditable. A spec that uses “good writing” as a requirement is not a spec. It is a wish. The model cannot deliver against a wish. The model can deliver against a binary.
The spec is the work. The spec is what makes the sentence mine. The model that writes the sentence is the contractor. The model that follows the spec is doing what a good contractor does. The model that ignores the spec is fired and rebid.
This sounds dramatic for a piece of writing. It is. Writing is a building. The spec is the plan. The model is the crew. I do not let the crew decide what the building is for.
The recursive audit, the gate#
The discipline is not delete-keep-rewrite. The discipline is audit-rewrite-audit.
After the model delivers a draft against the spec, I do not read it as a draft. I read it as a candidate for approval. Before I read it, I tell the model to audit its own output against the spec. Find every line that does not meet the standard. Rewrite each line. Audit again. Continue until the audit returns clean. Then deliver.
The model is good at this. The model is better at it than I am. The model can run the audit a hundred times in the time it would take me to read the draft once. The model knows the rubric. The model has no ego. The model has no fatigue. The model will not stop the audit on the fifth pass because it is bored. The model stops when the spec stops finding violations.
The audit instruction I use is short. Go through this draft line by line. For each line, check it against the spec. If any constraint is violated, name the constraint, name the line, and rewrite the line. Then audit again. Continue until there are no violations. Then deliver. That is the whole instruction. It is what I give the model every time.
I deliver to myself only what has passed the recursive audit. The audit is the gate. The gate is mine.
This is sharper than asking whether I chose the thought or accepted it. The question is: did the draft pass the audit? If yes, the work goes forward. If no, the model continues. The author is the one who wrote the spec and held the standard. The author is not the one who typed every word.
When I write a new essay, I give the model the spec and the topic. The model writes the draft. The model audits the draft against the spec. The model rewrites until the audit clears. I read what passes. I revise from there. The work is in the spec, not in the sentence. The work moved up one level. That is the architect’s move.
This is also how to read AI claims in the news. The layer that matters is the layer where the spec lives. The model does what the spec asks. The spec is the question. The model is the answer. The conversation is about the spec.
The fit is what I am responsible for, spec-by-spec, not sentence-by-sentence#
The model finishes thoughts I have started. The thought is mine because I started it. The completion is the model’s. The fit between the two is what I am responsible for.
I do not write sentence-by-sentence with the model anymore. I write spec-by-spec. The model writes the sentences within the spec. I audit the spec, not the sentence. The audit of the sentence happens in the model’s own work, against my spec, before I see it.
What this changed for me is the location of the author. The author used to be the person who typed the words. The author is now the person who set the standard the words have to meet. Both versions are real. The second one is the version that scales. The first one is what most writers are still doing.
If you write with AI and the writing does not sound like you, your spec is missing. Write the spec. Hand the model the spec. Tell the model to audit. The voice will come back. The voice was always in the spec.
This is also how I work with the engineers I hire. I do not write code for them. I write the requirements. I audit against the requirements. I approve when the work meets the standard. The model is the same. The model is the contractor. The contractor changes. The architect does not.
On a Tuesday morning, the spec is open on my desk. The essay topic is in my head. I write three sentences of opening, the way I want the essay to start. The model takes the three sentences and the spec and writes a draft. The model audits the draft against the spec. The model rewrites until the audit clears. The model delivers. I read what the model delivered. I revise the spec if the spec was the problem. I revise the sentences if the sentences were the problem. The work I did is what an architect does. The work the model did is what a crew does. The result is a building.
What the decision before the sentence was always pointing at, for me, is the decision before the spec. The spec is the place where I decide what I think. The model is the place where the deciding becomes the writing. The audit is the place where the writing has to match what I decided.
What the spec protects, voice, argument, standard#
The voice stays mine because the spec was mine.
The argument stays mine because the rubric was mine.
The standard stays mine because I wrote it down before the model touched the page.
The writing gets better. The model can audit against the spec more times than I can. The model is faster than I am at finding violations. The model is more patient than I am at the hundredth pass. The hundredth pass is the one most writers do not get to. The hundredth pass is the one the spec earns for free.
The reader gets a sentence that has passed an audit. That is more than most published writing on the internet. Most published writing has not passed any audit. Most published writing was not built against a spec. The reader cannot tell the difference at first. The reader can tell over thirty essays.
A reader who reads two essays on this site will not know how the work was made. A reader who reads ten will notice that the rhythms repeat in a way that does not repeat in most blogs. The rhythms repeat because the spec specifies them. The household metaphor is in every essay because the spec requires it. The close is the shortest sentence because the spec requires it. The voice does not drift because the spec does not drift.
The spec is the source of what the reader experiences as voice.
I will write a sentence tomorrow. I will hand it to the model. The model will rewrite it against the spec. The model will audit. The model will rewrite again. I will read what passes.
The sentence the model delivers will be a sentence I could have written. I will know it is mine because the spec was mine. I will not have typed every word. I will have written the standard every word had to meet.
The other half of this argument is the essay about what the model cannot do, no matter how good the spec. Together they are the working agreement I have with the model. The spec is what makes it useful. The limit is what makes it human.
The light is on at the desk. The spec file is open. The model is waiting.
Audit.
