Contents
While developing The Architecture of Nonfiction, I wrote this sentence:
Before
Clarity matters, but it is often a local request.
The sentence was grammatical. It sounded technical. It also failed to say what local request meant in a discussion of writing.
I reviewed the phrase, realised it needed work and asked my AI editor for help. My editor, Codex, replaced it.
After
Clarity matters, but we usually judge it one sentence or paragraph at a time.
The revision named the actual levels under review. It replaced a compact abstraction with the relationship the reader needed to understand.
That edit fixed one sentence. It did not prevent the same problem from appearing again.
Other phrases in the draft had the same shape. Related failure modes required the reader to determine which failures were related. Those instincts described skills that technical practitioners develop through training and experience as if they were innate. These traditions grouped prior art and scholarship into a category neither source belonged to. A sentence about what readers must carry used a familiar metaphor where the specific requirement was context they had to retain.
Each correction improved its sentence. Together, they revealed a durable editorial rule:
Name the actor, action, unit, or relationship instead of hiding it inside a neat-sounding abstraction.
That rule now lives in my author style guide. It is available when we draft a new article, revise an old one, review a newsletter, or ask a new AI session to edit a passage it has never seen before.
The style guide did not begin as a description of how I hoped to sound. It grew from decisions about how my writing should work.
One Edit Is Not Reusable Judgment
An editor can replace local request without explaining the decision. The sentence improves, the document moves forward, and the reasoning disappears into the completed edit.
That is sufficient when the problem is isolated. It is wasteful when the correction expresses a preference that should apply again.
Software teams encounter the same difference between a patch and a standard. A senior engineer can fix one defect during a code review. The organization gains more when the underlying decision is documented as a rule future work should follow, supported by a code example or an automated check.
Authorial judgment deserves the same treatment.
An edit becomes reusable when it records more than the replacement text. It should preserve the behavior the author rejected, the reason for rejecting it, the contexts where the decision applies, and at least one acceptable implementation. Without that information, the next collaborator must rediscover the decision from another round of feedback.
The guide carries decisions beyond the current sentence. It should not freeze every preference forever.
The Model Had Project Context, Not Author Context
The parent essay did not lack context. We had source notes, an outline, a working draft, prior articles, research, and a long editorial conversation. Those artifacts described the subject, argument, evidence, structure, and current state of the work.
They did not automatically describe me.
The project context could tell an AI editor that a section concerned Single Responsibility. It could not tell the editor that I reserve claim for factual, predictive, or externally verifiable assertions. The outline could identify the intended audience. It could not show that I dislike prose as the default word for every article, document, draft, paragraph, passage, and text. A source note could establish what a study found. It could not decide whether a parallel phrase performed useful rhetorical work or merely sounded clever.
Those decisions were available while they remained in the working conversation. A new session would not inherit them reliably. Neither would a human editor joining the project later.
Conversation history is useful memory. It is a poor source of truth. Important instructions become mixed with abandoned ideas, temporary decisions, resolved disagreements, and details that mattered only to one paragraph.
A maintained style guide separates durable author context from the history that produced it.
Style Is More Than Tone
Ask an AI to write in a clear, professional, punchy, and authentic voice, and you have supplied aspirations. You have not explained what those words mean for this author, this audience, or this publication.
Punchy might produce a wall of one-line paragraphs. Professional might replace plain language with corporate abstractions. Clear might remove a necessary qualification. Authentic might imitate the surface features of a few examples while missing the judgment behind them.
An author-specific style guide has a different responsibility from the other guidance surrounding a piece:
| Artifact | What it defines |
|---|---|
| Brand guide | Public identity, visual presentation, and broad messaging |
| House style | Publication-wide usage and consistency |
| Project instructions | Repository, workflow, and deliverable requirements |
| Task prompt | The current assignment |
| Sources and draft | Evidence and the current implementation |
| Author style guide | The individual author’s recurring editorial judgment |
These artifacts can overlap. A publication may forbid em dashes as a house rule, while an author avoids them as a personal convention. A repository may require a Receipts section, while the author’s evidence standard determines how facts, first-party observations, inferences, and predictions are distinguished inside it.
The useful distinction is responsibility. The author style guide records how this person makes recurring writing decisions. That includes voice, but it also includes audience assumptions, evidence boundaries, terminology, rhetorical tolerances, cadence by channel, and patterns that deserve extra scrutiny.
Runtime Context Changes the Work in Front of the Model
In software, runtime context is information available while a system performs its current work. The executable may remain the same while inputs, configuration, permissions, or environmental state change its behavior.
The comparison is useful for AI-assisted writing when we keep its boundary clear. A general model brings learned language capability and default patterns. The current prompt, sources, draft, examples, and instructions supply information for the present task. An author style guide is one maintained part of that context.
| AI-assisted writing element | Role |
|---|---|
| Model | General language capability and learned defaults |
| Task prompt | Current assignment |
| Sources and draft | Working inputs |
| Author style guide | Reusable editorial constraints, preferences, and examples |
| Author or editor | Final authority over the result |
Loading a style guide does not retrain the model or change its underlying weights. Natural-language instructions also do not execute with compiler-enforced semantics. The model can ignore a rule, misread its scope, apply it mechanically, or follow two instructions that conflict.
The guide gives the collaborator more of the information needed to make an editorial decision. It does not guarantee the result.
That limitation does not make the practice weak. Human collaborators also need context, examples, and correction. A style guide gives both human and AI editors a maintained starting point instead of asking each one to infer the author’s standards from the current draft.
Research on academic revision points in the same general direction. Nuo Chen and colleagues designed XtraGPT around criteria-guided intent alignment and context-aware modeling rather than surface-only prompting. Their evaluated system used 140,000 instruction-response pairs derived from section-level revisions across 7,000 research papers, and it outperformed same-scale baselines in their experiments.
XtraGPT is a specialized academic-revision system. It does not evaluate author style guides, and it does not establish that a Markdown file is the best implementation for every workflow. It does support the underlying design choice: revision improves when the system receives explicit criteria and enough surrounding context to understand the work beyond one sentence.
General Models Do Not Begin as You
AI writing assistance can influence more than grammar.
Dhruv Agarwal, Mor Naaman, and Aditya Vashistha ran a controlled study with 118 participants from India and the United States completing culturally grounded writing tasks with or without AI suggestions. Their analysis found that AI suggestions increased similarity within both groups. It also found cross-cultural convergence, with Indian participants’ writing moving more toward American styles than the reverse.
That study examined a specific cross-cultural setting and a particular form of inline assistance. It does not prove that every AI-assisted document becomes uniform. It does establish that suggestions can influence how people write, including features more subtle than corrected grammar or removed cultural references.
Tuhin Chakrabarty, Philippe Laban, and Chien-Sheng Wu approached the problem through professional editing. Writers in their study identified recurring undesirable patterns in LLM-generated creative passages, including clichés and unnecessary exposition. The researchers assembled 1,057 edited paragraphs, and their preference evaluation found that experts largely preferred passages edited by other experts.
That research studied creative domains, not technical essays. Its categories are not my style guide. The important result here is narrower: general models produce recurring writing patterns, and professional judgment still changes which implementations experienced readers prefer.
An author-specific guide makes one person’s judgment available before generic defaults accumulate across an entire draft.
Encode Decisions, Not Vibes
A useful rule tells the collaborator what to notice and what to do with it. It also explains why the rule exists, where it applies, and when an exception may be justified.
Compare these instructions:
Vague preference
Make the writing sound more natural.
Maintained editorial rule
Prefer a concrete noun or verb over a leadership abstraction. Name the actor, action, unit, or relationship when a compact phrase would force the reader to infer it. Preserve technical terms when they supply a defined diagnostic.
The second version does not prescribe one sentence shape. It identifies a failure, gives a corrective action, and protects a legitimate exception.
The minimum useful author guide covers more than word choice:
- Audience and outcome: Who the author expects to reach and what the writing should help them understand, decide, or do.
- Voice and positioning: How authority, uncertainty, experience, disagreement, and direct address should appear.
- Format and cadence: How long-form explanation differs from a LinkedIn post, newsletter, white paper, or internal standard.
- Evidence and attribution: What requires a source, how first-party observations are labeled, and how facts are separated from inference.
- Terminology: Which words are preferred, which have precise meanings, and which should be used sparingly.
- Suspect patterns: Constructions that deserve review because the author or model overuses them.
- Examples: Before-and-after passages, positive examples, negative examples, and the reasons they differ.
- Authority: What happens when the guide conflicts with evidence, project requirements, or the present document’s purpose.
Some entries can be mechanical. My writing does not use em dashes, and every list of three or more items takes an Oxford comma. Other entries require judgment. An isolated sentence may create emphasis, but a long-form article should not become a ladder of one-line paragraphs. Repetition can build rhythm, but it should not reveal a template applied without regard for meaning.
A guide becomes more useful when it distinguishes mandatory rules from defaults that permit earned exceptions.
Build the Guide From Edit History
The strongest entries in my guide came from finished work, editorial disagreement, and repeated revision. We did not invent them by selecting adjectives from a branding exercise.
The parent essay produced several examples:
| Observed edit | Durable rule | Boundary |
|---|---|---|
| Replaced local request and similar abstractions | Name the actor, action, unit, or relationship | Keep technical language when it supplies a defined diagnostic |
| Restricted generic uses of claim | Choose argument, observation, judgment, hypothesis, or claim by evidentiary function | Keep claim for factual, predictive, or externally verifiable assertions |
| Reduced 64 uses of prose | Name the actual object: writing, article, document, draft, paragraph, passage, or text | Keep prose when contrasting written language with code or naming a defined concept |
| Reserved test after repeated overuse | Use test for a defined procedure with an observable result | Use criterion, standard, review question, or verification step elsewhere |
| Protected “from remembrance to responsibility, and from responsibility to resolve” | Semantic specificity should not flatten rhetoric that performs useful work | Judge the effect against the passage’s responsibility |
This history matters because rules written in the abstract often describe an imagined author. Revision history records the decisions the author actually makes when a sentence, paragraph, or section is in front of them.
The negative examples reveal recurring failure shapes. The accepted revisions show how the author resolves them. The rationale prevents a future editor from copying the surface change while missing its purpose.
Not every correction should be promoted. If I replace one word because it sounds awkward in one sentence, that preference may end there. If the same issue appears several times, exposes a stable distinction, or protects an important evidence boundary, it deserves a maintained entry.
The guide should also change when the author changes. A rule may become more precise, gain an exception, split by publication channel, or disappear when it no longer represents the work. Versioning the guide makes that development visible without pretending the first description was permanent.
Examples Make a Rule Usable
Rules such as be concise, avoid jargon, or write with authority leave most of the difficult decision to the collaborator. Examples provide the missing boundary.
A negative example shows the shape of the failure. A positive example shows one acceptable implementation. The explanation between them identifies what the editor should preserve when neither exact sentence fits the next draft.
This is why the local request revision contributes more than the instruction to use specific language. The pair shows that a grammatically valid, technically flavored phrase can still be too vague. The explanation identifies the missing unit of analysis. The broader rule then transfers that decision to different wording.
Examples turn an adjective into an editorial decision another collaborator can examine and apply.
Authority Must Be Explicit
A style guide should make writing more consistent. It should not make the author consistently wrong.
The source evidence retains authority over the conclusion. Legal, safety, accessibility, and publication requirements retain authority over stylistic preference. The current document’s purpose can justify a different cadence or vocabulary from the author’s usual defaults.
A practical order looks like this:
| Priority | Authority |
|---|---|
| 1 | Facts, source evidence, legal requirements, safety constraints, and accessibility needs |
| 2 | Mandatory repository, organization, or publication-wide requirements |
| 3 | The current task’s explicit requirements and the document’s purpose, audience, and deliverable requirements |
| 4 | Author-specific defaults and preferences |
| 5 | Model preferences and generic writing conventions |
The exact middle order may change by organization. A task-specific instruction may override a general default, but it cannot cancel a mandatory publication requirement. The important decision is to define the order before two instructions conflict.
My style guide asks writers to distinguish facts, first-party observations, inferences, and predictions. That rule helps preserve evidentiary boundaries. It cannot require a favored conclusion after a new source contradicts it.
A style guide can govern how evidence is presented. It cannot change what the evidence supports.
Load the Smallest Relevant Guide
An author guide can grow to cover long-form cadence, newsletter structure, LinkedIn constraints, evidence standards, recurring terminology, and the treatment of work anecdotes. A single task rarely needs every section.
Loading irrelevant guidance consumes attention and creates opportunities for conflict. A LinkedIn revision needs the author’s voice and short-form constraints. A research-heavy companion essay needs the evidence rules, long-form cadence, terminology, and Receipts conventions. It does not need image-prompt guidance or a newsletter call to action.
Loading means making the selected guidance available to the current AI session. You might place it in project instructions, attach the guide as a context file, or paste the relevant excerpt into the task. The mechanism makes the guidance available. It still does not guarantee that the model will apply it correctly.
Your AI Instructions File Should Be a Router, Not a Novel explains how a short entry point can direct an AI collaborator toward the standards relevant to the current work. The relationship is simple: the router decides which guidance to load; the author style guide defines one category of guidance.
Selection prevents a useful guide from becoming another brain dump.
The Author Still Owns the Result
An AI editor can follow a valid rule badly.
It can replace every abstraction until the writing becomes tedious. It can treat a warning about one-line paragraphs as a prohibition against emphasis. It can reserve claim correctly while choosing a less accurate synonym. It can remove parallel phrasing that carries an argument because repetition appears on the list of patterns to inspect.
That last failure nearly happened in the parent essay. A review challenged the phrase “from consecration to conviction” as decorative alliteration. The rhetorical parallel was intentional, but the wording did not accurately follow Lincoln’s movement from remembering sacrifice, to accepting responsibility for unfinished work, to resolving that the sacrifice will not be in vain. We revised it to “from remembrance to responsibility, and from responsibility to resolve.”
The right decision was neither “keep the author’s words” nor “remove the rhetoric.” It was to preserve the function while improving the implementation.
No static rule can make that judgment in every context. The guide can identify the author’s preferences, supply examples, and establish review questions. The author still decides whether a rule applies, whether an exception is earned, and whether the result fulfills the document’s purpose.
Portable judgment still needs a judge.
Turn Corrections Into Context
You do not need to write a complete style guide before using AI. Start with the evidence your writing already provides.
- Collect representative work. Use published pieces, drafts, revision history, editorial comments, and examples you still consider successful.
- Find repeated decisions. Look for corrections you make often and choices you repeatedly protect during review.
- State the durable rule. Name the behavior, the preferred response, the reason, and the scope.
- Add examples. Pair rejected and accepted implementations, then explain the difference.
- Separate contexts. Distinguish universal requirements from defaults for long-form, social, internal, technical, persuasive, or research-heavy work.
- Declare authority. Explain what wins when the guide conflicts with evidence, the task prompt, project instructions, publication rules, or the present purpose.
- Load the relevant sections. Give each collaborator the guidance required for the current work instead of every rule ever written.
- Maintain the guide. Promote new decisions when they recur or protect an important boundary. Revise old rules when they stop representing your judgment.
Then use the guide during drafting and review. Ask the AI collaborator to identify which rules it applied, where the draft creates a conflict, and which decisions still require the author.
The goal is not to make every sentence look the same. It is to stop rediscovering the same editorial knowledge sentence by sentence.
A useful style guide gives the next collaborator more than a description of your voice. It gives them the decisions that created it.
Receipts
- First-party revision history: The local request example, the semantic-specificity rule, the reserved uses of claim and test, the prose terminology review, and the protected rhetorical parallel come from the documented development of The Architecture of Nonfiction. They demonstrate how this author’s maintained guide was built. They are not presented as universal preferences.
- Cross-cultural writing influence: Dhruv Agarwal, Mor Naaman, and Aditya Vashistha’s “AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances” reports a controlled experiment with 118 Indian and American participants. The study found within-group homogenization under AI assistance and stronger movement of Indian writing toward American styles than the reverse in the tasks studied.
- Professional editing of LLM writing: Tuhin Chakrabarty, Philippe Laban, and Chien-Sheng Wu’s “Can AI writing be salvaged?” reports professional writers’ agreement on recurring undesirable patterns in LLM-generated creative passages, a corpus of 1,057 expert-edited paragraphs, and a preference evaluation in which experts largely preferred expert-edited text. The study does not evaluate technical essays or author-specific style guides.
- Criteria-guided, context-aware revision: Nuo Chen and colleagues’ “XtraGPT: Context-Aware and Controllable Academic Paper Revision via Human-AI Collaboration” describes a specialized academic-revision framework centered on criteria-guided intent alignment and context-aware modeling. The authors evaluated models trained with 140,000 instruction-response pairs derived from 7,000 research papers. Applying that design direction through an author-specific style guide is our implementation, not a result the study tested.