Branch

Modeling the Author

Human Duke

Everyone agrees that LLMs are pretty terrible at Essay writing. They use a repetitive and easy to spot style. They don't usually say anything interesting either.

We've seen that agent harnesses like Claude Code and Codex have given software engineers super powers, but there's nothing like this yet for essay writers (or writers in general).

What sort of tooling would give writers super powers?

How does it look different from the work software engineers do?

How can writers retain their unique style and voice? Or how can the tooling help them develop one?

AI · author-edited OpenAI

The equivalent of Claude Code for writers is probably not “a model that writes the essay.” It is a research, argument, and revision environment in which the writer remains the source of judgment.

Coding agents became useful when they gained a harness: access to the repository, search, tests, version control, execution, and a loop for checking their work. A writing agent needs analogous machinery—but prose has no compiler, and “good” is not a single measurable state.

What a writer’s harness would contain

1. A living model of the project

Not just a chat transcript or a folder of documents. The system should maintain:

  • the intended reader
  • the central question
  • candidate theses
  • claims and supporting evidence
  • objections and unresolved tensions
  • sources, quotations, and provenance
  • definitions of important terms
  • examples, scenes, and anecdotes
  • structural outlines
  • stylistic constraints
  • decisions the writer has made and rejected

This would be closer to a graph than a document. An essay is merely one rendering of that graph. WRITER.md

2. An argument debugger

This may be the closest analogue to tests and static analysis.

It could identify:

  • hidden assumptions
  • equivocation between two meanings of a term
  • causal claims supported only by correlation
  • conclusions stronger than the evidence permits
  • unaddressed obvious objections
  • false dichotomies
  • sections that are individually plausible but do not advance the thesis
  • abstractions that need examples
  • examples that are vivid but unrepresentative

It should also be able to “run” the argument against several readers: an expert, a hostile critic, an intelligent newcomer, or someone directly affected by the topic.

This is not an automatic truth machine. It is a way to generate pressure against the writer’s reasoning. Logical Fallacy Linter

3. Structural editing

Most writing assistance operates one sentence at a time, which is often the wrong level.

A useful system would let the writer manipulate the essay at several scales:

  • argument: claim, evidence, objection, reply
  • section: purpose, tension, payoff
  • paragraph: assertion, development, transition
  • sentence: rhythm, syntax, diction
  • word: precision and connotation

It could display a “reverse outline” generated from the actual draft, then compare that with the intended outline. It might label each paragraph by function and expose six consecutive paragraphs all doing background exposition.

It should support operations like:

  • “Show me three structures for this material without drafting them.”
  • “Move the central complication earlier.”
  • “What can be removed if this must lose 30% of its length?”
  • “Find the point at which the essay stops asking a question and starts pretending it has answered it.”
  • “Preserve the conclusion, but propose a route to it that does not depend on chronological organization.” Structural Editing UX

4. Revision instruments, not a rewrite button

A generic “improve this” button almost inevitably pulls prose toward the model’s average style.

Better tools would expose specific transformations:

  • compress repetition
  • replace abstract claims with concrete instances
  • vary sentence architecture
  • make uncertainty explicit
  • separate observation from inference
  • remove throat-clearing
  • identify clichés and inherited metaphors
  • locate unexplained jumps
  • estimate where reader attention is likely to drop
  • compare two revisions and describe what each gains and loses

The key is that suggestions should arrive as diagnoses, alternatives, or patches, not as a silently regenerated document.

How writing differs from software engineering

There are important analogies:

Software Writing
Repository Notes, sources, drafts, archive
Dependency graph Claim/evidence graph
Compiler errors Grammar and consistency failures
Unit tests Factual and citation checks
Integration tests Argument coherence
Runtime profiling Reader-response and pacing analysis
Version control Draft history and decision log
Code review Editorial and adversarial review

But the differences matter more.

Writing has no complete specification

In software, the desired behavior can often be stated externally. In an essay, discovering what the writer really thinks is part of the work. The specification changes because composition changes the author. Modeling the Author You are here

This branch begins here Modeling the Author
Human Duke

Over time the decisions a writer makes need to be incorporated into a "Theory of Mind" about the writer. The harness needs to update itself and it's instructions as the writer learns new things about themselves and their topic.

AI OpenAI

Exactly. The harness needs a dynamic model of the writer, not merely a static style guide. But “Theory of Mind” should be treated as a set of revisable hypotheses, not a definitive description of the person.

Otherwise the system will freeze contingent habits into identity:

“You usually prefer short openings” becomes “You are a writer who must use short openings.”

That would prevent development rather than support it.

A good system would maintain two coupled models:

1. A model of the writer

This records patterns such as:

  • recurring intellectual interests
  • aesthetic preferences
  • characteristic forms of argument
  • tolerance for ambiguity
  • preferred relationship with the reader
  • recurring strengths and weaknesses
  • phrases, rhythms, and syntactic structures they favor
  • kinds of evidence they find convincing
  • feedback they repeatedly accept or reject
  • tendencies under uncertainty—for example, overqualification or premature certainty
  • ambitions they have stated but have not yet realized in their prose

Crucially, these should have provenance and confidence:

Hypothesis: The writer prefers to begin with a concrete anomaly rather than a general thesis.
Evidence: Accepted this change in essays A and C; explicitly requested it in B.
Confidence: Moderate.
Possible exception: Technical explanatory pieces.

The model should distinguish at least four kinds of information:

  1. Explicit commitments — “I don’t want to sound omniscient.”
  2. Observed preferences — the writer repeatedly deletes rhetorical questions.
  3. Project-specific decisions — this essay should avoid autobiography.
  4. System hypotheses — the writer may use abstraction to avoid making a vulnerable claim.

Only the first and perhaps second categories should strongly affect future behavior. Hypotheses should be offered back for reflection, not silently installed as rules.

2. A model of the inquiry

The topic model should evolve alongside the writer model:

  • what the writer initially believed
  • what evidence changed their mind
  • which distinctions became important
  • which questions remain unresolved
  • which sources are authoritative, disputed, or obsolete
  • where the draft is more certain than the writer actually is
  • which attractive claims were abandoned, and why
  • how the central question itself has changed

That last item is especially important. Often the deepest progress in an essay is not finding a better answer but realizing that the original question was malformed.

The system might say:

You began by asking why remote work reduces creativity. Your notes now suggest a narrower and different question: which kinds of creative coordination depend on shared context, and which merely depend on good documentation? Should the essay’s framing change?

That is much more useful than generating another polished paragraph.

The harness should learn from decisions, not just final prose

Final drafts are poor training data for understanding a writer. They conceal the process that produced them. The valuable information is in the sequence:

  • The model suggested three openings.
  • The writer rejected two.
  • They combined the third with an earlier anecdote.
  • They explained that the rejected versions announced the lesson too soon.
  • Later, they made the same choice in the conclusion.

This reveals a principle: the writer values delayed interpretation, allowing an example to acquire meaning before explaining it.

So the basic unit of learning should be the decision event:

  • what alternatives were available
  • what the writer selected
  • what they changed manually
  • what reason they gave, if any
  • whether the decision survived later revision
  • whether it appears to be local or generalizable

A writing harness would therefore resemble version control plus an editorial lab notebook. It would preserve not only drafts but the rationale behind revisions.

Self-updating instructions need a promotion process

The system should not rewrite its permanent instructions after every interaction. It needs levels of memory:

Working context

Temporary assumptions for the current passage or session.

Project memory

Decisions that govern one essay or book:

  • terminology
  • audience
  • source policy
  • narrative distance
  • structural commitments
  • unresolved questions

Writer model

Cross-project tendencies and explicit preferences.

Constitutional commitments

A small set of principles the writer deliberately adopts, such as:

  • Never invent quotations or citations.
  • Distinguish observation from inference.
  • Do not resolve genuine ambiguity merely to produce a stronger ending.
  • Prefer exactness over apparent authority.

Movement between levels should require evidence or confirmation. After several repeated decisions, the harness could ask:

Across four projects, you have rejected edits that replace qualification with certainty. Should I adopt “preserve epistemic calibration” as a standing instruction, or is this topic-dependent?

This makes learning visible and reversible.

It should also model aspiration, not merely imitation

If the harness learns only from previous work, it becomes a machine for reproducing the writer’s past. Writers often want to become capable of something they cannot yet do.

The writer model therefore needs a distinction between:

  • current voice
  • declared values
  • developmental goals
  • experimental modes

For example:

Current tendency: explanatory, highly qualified, low use of scene.
Stated goal: greater narrative momentum.
Constraint: do not manufacture drama or simplify the argument.
Experiment: open the next three pieces with an unresolved incident.

The harness can then help the writer stretch without replacing their voice with a generic “better” one.

This also suggests temporary style branches, analogous to Git branches:

  • “More severe and compressed”
  • “More intimate, without becoming confessional”
  • “Assume a skeptical expert reader”
  • “Try a structure built around scenes”
  • “Write under the influence of Didion’s movement between detail and claim, without imitating her diction”

The writer can experiment, compare results, and selectively merge what works into their evolving practice.

The model should be inspectable and contestable

A hidden personalization model would be dangerous and creatively narrowing. The writer should be able to ask:

  • What do you currently believe about my writing?
  • Which observations support that belief?
  • What instructions have you inferred?
  • Which old preferences are shaping this suggestion?
  • Where do my stated goals conflict with my demonstrated choices?
  • What have I changed my mind about?
  • Forget this preference.
  • Treat this as an experiment, not part of my permanent profile.

It should periodically surface contradictions:

You say you want concise prose, but you consistently preserve long sentences when they accumulate examples. Perhaps your actual preference is not brevity but structural clarity.

That is a genuinely useful theory of mind: not a flattering profile, but a mirror the writer can argue with.

The deepest version of this harness is therefore not an automated ghostwriter. It is a long-term intellectual collaborator with memory—one that remembers how the writer’s thinking developed, notices when old instructions no longer fit, and helps distinguish a durable principle from a habit they may be ready to outgrow.

Explore conversation