How to write with AIWriting as a compilation process, from scattered notes to a clear argument and readable prose.

I have more material than I can turn into essays. I write journals and talk through ideas on walks. By the time I sit down to write, I have fragments from several different conversations about the same thing.

AI is useful here because it can help me recover those fragments and work out how they belong together. Sometimes a raw dump and a request for a polished essay are enough. When the result misses, though, another rewrite can improve individual sentences without getting much closer to the piece I wanted.

I want writing that communicates what I mean and doesn't make the reader work through an unpleasant synthetic voice to receive it. To make that work with agents, I needed a way to specify the decisions behind the sentences.

Substance and voice need different edits

When a draft feels wrong, I ask whether it gives the reader something they didn't have before and whether its points follow from one another. I also ask whether the prose is pleasant to read. I may agree with the argument and still dislike reading sentences that feel airy, repetitive or assembled from familiar phrases.

Those questions lead to different edits, but the dimensions affect each other. Missing reasoning can make a sentence generic, and bad phrasing can obscure a precise thought. Separating the questions helps me avoid treating every problem as a request to “make it sound more human.”

I call the things a piece gives its reader substantives. A useful fact qualifies, but so does a nuanced argument whose value only becomes clear after several paragraphs. A worked explanation can be the main reason to read it.

For Thinking in Rhythms, I wanted to unpack the promise of agents working around the clock. Work waits for incoming information, another task to finish, or a human decision. Because work has dependencies, removing the need for sleep doesn't remove every reason to wait. The right schedule depends on what can become actionable between runs.

An outline containing “autonomy,” “triggers” and “human rhythms” wouldn't preserve that argument. Those headings name topics. A causal spine records why one point follows from another, which is the part I needed the reader to understand. If that relationship is missing, another voice pass won't supply it reliably.

AI-generated content sometimes reminds me of scar tissue. It is functional and continuous, but it lacks the texture of the thing it replaces. I feel that when every sentence resolves into a familiar shape, as though a small set of phrases were doing the work of a much more complicated understanding. The text could move to another topic with a few nouns changed. It doesn't seem to have encountered the awkward case that would force the writer to qualify the claim or choose a more exact word.

Simple language can carry a complicated model. “The task was waiting for my decision” is plain and precise. “The task had reached a point where human input was required for further progress to be possible” makes me work harder to receive the same information.

Other cases need more explanation. A paragraph about running agents faster needs to distinguish finishing earlier from reaching the next useful step earlier. Removing that distinction might make the paragraph shorter while making the thought less accurate.

I associate good writing with the pressure of those decisions. The writer's understanding includes cases where an obvious claim fails, so a qualification has a reason to be there. Their sense of the reader changes which example they choose and which steps need explaining.

I want the specific thought to determine the prose. Adding quirks to disguise AI involvement would give me another layer of performance to remove. My suspicion is that much of the complaint about AI writing comes from the experience of reading something bad. The prose promises an insight and delivers a familiar conclusion. Whether AI helped produce a piece matters less to me than whether the piece was worth reading.

Writing is a compilation process

A sentence belongs in an article because of what its paragraph needs to explain. The paragraph's job depends on where the argument has reached, which depends on the structure of the piece. Even an elegant sentence can be wrong if those earlier decisions are wrong.

Following those dependencies backward brings me to the thing I want to communicate. The spark determines which questions I research. What I learn helps me choose a form. The form and material constrain an outline, which gives the sections and paragraphs their jobs. The sentences make those decisions visible.

This is why I think of writing as a compilation process. An intention becomes an account of what I mean, then a structure for explaining it, then prose someone can read. Each representation makes decisions that the next one relies on.

Stage	What it settles
Spark	What I want to share, and why it might matter.
Research	Which material belongs, what it supports, and what remains unclear.
Form	Whether the piece should explain, argue, teach a process or do something else.
Causal spine and outline	What the reader needs to understand first, and why each point follows.
Headings and paragraph beats	What each section and paragraph must accomplish.
Sentences	How each paragraph expresses its particular thought.
Stylistic choices	The words, rhythm and transitions through which the reader receives it.

These decisions still have to be made when I ask for an essay in one prompt. The model fills in whatever I have left open. Feedback such as “more insightful” or “less AI” asks it to revise the result without identifying which decisions failed or which ones should stay fixed. The next version might improve the voice while quietly changing the point.

That explained my frustration with stylistic passes. I was trying to settle individual expressions while the claims and structure were still changing. Editing only the sentences was like patching binary bytes while the source program kept changing. Unless the reason for a correction reached the source, the next compilation could lose it.

To delegate a stage, I need to specify what its output must preserve. Suppose a paragraph must explain why an agent that finishes earlier can still wait until the same human review. The model can choose an example and find the wording. A paragraph that merely recommends running agents more often has failed the assignment, however well it reads.

I call that specification a macrostate. It describes conditions that many different outputs can satisfy. Each stage reduces the degrees of freedom for what comes next, while leaving room for choices that don't change the intended result. Once a paragraph's meaning and role are settled, I can leave much of its expression to the model.

The intermediate decisions make the conditions I started with specific enough to delegate: the piece has something worth reading, and the prose isn't painful to read. I don't need one uniquely correct essay. I need to know which differences are harmless and which ones violate what I wanted to communicate.

From a spark to a working brief

“What does it mean for an agent to work for you 24/7?” was the useful question for Rhythms. It gave the research a purpose. I was looking for material that could help resolve that question, rather than everything I had ever said about agents.

An archive search can help find a spark, too. Several notes might return to the same unresolved problem. I still need to choose what I want to communicate, because the most frequently mentioned topic isn't necessarily the thought worth developing. When a phrase already expresses the thought well, I keep it with the brief so the surrounding paragraphs can grow around it. Adding the phrase to a finished essay later risks making it fit a structure that was built for another point.

A walking note often assumes context I had in my head. It may contain the conclusion without the event that prompted it, or a complaint without the distinction that would make it useful to someone else. For the rhythms question, relevant material included why teams meet at regular intervals, where work accumulates and what a person actually decides when an agent needs help. A note about a model's reasoning ability might share the same vocabulary while contributing little to that explanation.

A collection of matching quotations still leaves the writing problem unsolved. I want an account of what those quotations mean together, where they disagree and what remains unclear. AI can propose connections I hadn't articulated, but I need to distinguish its interpretation from something I actually said. An explanation that sounds coherent can still assign me a position I don't hold, or turn a tentative idea into a conviction.

My working brief contains:

• The thought I want the reader to receive, and what prompted it.
• What the reader probably believes already, including the confusion the piece should resolve.
• The relevant evidence and what it supports.
• The claims and distinctions that must survive, including anything still uncertain.
• Any exact phrases worth preserving.

That brief is easier to disagree with than a polished essay. I can point to an interpretation that is wrong before several paragraphs depend on it.

Once the research has clarified the thought, I choose a form. A guide needs to leave the reader able to do something. A mental-model essay needs to develop a way of seeing a problem and show what that perspective changes. I use existing articles as references for how those explanations develop. A useful template might introduce a familiar problem, follow a concrete case until the usual explanation fails, then derive a better model from that failure. It tells the writer what each part must accomplish without borrowing the reference article's claims or anecdotes.

Within that form, I work out the causal spine and each paragraph's job. If the outline contains a claim and a conclusion but no explanation between them, I can address that before generating the prose.

Saving decisions and revising the draft

Chat is useful for challenging an interpretation or explaining why something bothers me. But the current understanding needs a home outside the sequence of attempts, so the next request doesn't have to infer which parts still govern the draft.

I keep one folder per essay in ~/projects/writing/essays/. BRIEF.md holds what the piece is trying to communicate. The pad/ folder holds research and the causal outline. Numbered drafts preserve earlier versions, and REVISIONS.md records what changed and why. My journals stay in my personal archive.

During substantive review, I ask whether the piece says what I meant and supplies the steps a reader needs. A comment such as “make this more insightful” gives the next attempt too much to guess. “This explains how to schedule a run, but the reader still doesn't know what could become actionable between runs” identifies a particular hole in the explanation. The correction belongs in the outline as well as the passage.

Once a section works, I keep it. A request to improve the ending shouldn't silently regenerate an opening I have accepted. If drafting exposes a contradiction, I return to the underlying claim and revise the paragraphs that depend on it. The saved decisions tell me where the change belongs.

Voice work after the substance holds together

I give the writer voice references early, then use a focused voice pass once the substance holds together. Unslop skills and filters can flag stock transitions, excessive symmetry or sentences that announce an insight instead of stating it.

My shared prose skill lives in my agent skills folder. structure.md governs how an explanation develops, register.md records patterns I keep rejecting, and examples.md shows better writing. A new session can read those files and the essay's brief before making a change.

Some rules are blunt. I don't use em dashes in these essays. I ask for complete sentences instead of fragments such as “Same craft, different physics.” I cut “Here's the core idea” when the next sentence can just state the idea, and flag repeated “it's not X, it's Y” constructions.

Other rules protect the explanation. Name who does what. Give “this” and “that” clear referents. Keep the steps that make a conclusion follow, even when they take another paragraph. A list of prohibitions can remove obvious tells while leaving the writer unsure what good prose should do, so I also keep positive examples of patient sentences that carry the reasoning.

When an entire section has the wrong rhythm, I would stop repairing it sentence by sentence. A fresh pass can use the settled argument and a positive voice example, with the selected phrases preserved, to produce that section again. Sections I have already accepted stay unchanged.

For this kind of writing, I find Opus 5 basically unusable. The prose is often so airy that I have to stop and work out what a sentence is saying. GPT 5.6 Sol has been much easier for me to steer toward the voice I want. These are my preferences from using them, and I expect them to change as the models improve.

Passing a style filter doesn't establish that a conclusion follows. I still have to read the explanation. If polishing removes the qualification that made a claim accurate, it has damaged the piece while making the sentences smoother.

Turning months of journals into an essay involves more unresolved decisions than turning a settled thought into a short message. If a direct request already gives me writing I would use, there is no reason to manufacture a larger procedure around it.

I expect better models to take on more of this work. What I want to retain is the spark and the understanding of what the piece should give its reader. Those let me judge the result, explain what missed, and decide when the writing is good enough to share.