--- schema_version: 1 type: "essay" id: "essay-co-writing-with-language-models" slug: "co-writing-with-language-models" title: "How I use AI models when I write" summary: "A practical method for using AI to move from thought to draft without losing the argument along the way." status: "published" visibility: "public" indexing: "index" categories: - "creativity" - "ai" - "thinking" tags: - "co-writing" - "writing process" - "voice" - "language models" created: "2026-03-19" updated: "2026-07-29" published: "2026-07-23" canonical_path: "/writing/co-writing-with-language-models" canonical_url: "https://rahilchamola.com/writing/co-writing-with-language-models" preferred_citation: "Chamola, Rahil. “How I use AI models when I write.” Curated Randomness, 2026." reading_minutes: 7 authors: - id: "rahil-chamola" name: "Rahil Chamola" role: "author" legacy_slugs: - "rules-for-co-writing-with-claude" - "co-writing-with-claude" series: id: "building-with-models" title: "Building with models" part: 3 related_slugs: - "10-things-ai-has-taught-me" citations: [] --- # How I use AI models when I write Writing with AI is useful to me because it makes ideas easier to examine. A half-formed thought can become an outline, a counterargument, or a draft while I still have the energy to work on it. However, the speed of that change also makes it easier to miss when an AI model has replaced a difficult part of the thought with something smoother and less precise. I do not think authorship depends on writing every sentence alone. It depends on knowing which claims are mine, which language came from elsewhere, what the evidence supports, and why the final structure says what I intend it to say. These rules are how I keep that responsibility visible. ## 1. Start with material only you can supply Before asking for prose, I try to provide the observation, question, tension, or view that made me want to write. I often do this by speaking because a transcript preserves the order in which the thought develops, including the places where I correct myself or remain unsure. An AI model may clean the transcript, identify its central claims, or ask where an example is missing. It should not invent the experience, supply a stronger belief than I expressed, or resolve a contradiction simply because finished prose usually sounds more settled. ## 2. Separate belief, evidence, and open questions What I believe, what I can demonstrate, and what I am still trying to understand are different parts of an article. Treating them as one category encourages confidence the evidence may not deserve. I therefore mark the distinction before drafting. A factual statement needs support. An interpretation should be presented as an interpretation. A live question may remain open if choosing an answer would be premature. An AI model can help me test each part, but it should not turn uncertainty into a thesis on my behalf. ## 3. Provide relevant context, not complete history For most articles, useful context is quite small: the intended reader, the current view, the evidence, a serious objection, a few lines whose language I want to preserve, and the limits of the piece. Adding years of private material does not necessarily produce better writing. It can mix past and present views, pull irrelevant experiences into the draft, and encourage an AI model to imitate a broad idea of me rather than work with the specific argument in front of it. I would rather choose the exact material that belongs in the article and keep everything else outside it. ## 4. Map the meaning before changing the prose A rewrite can preserve the main point while losing the structure that made the point defensible. It may remove a qualification, replace a precise term with a friendlier but weaker one, reduce two competing explanations to one, or split a sentence in a way that hides how its clauses depend on one another. For serious revisions, I first list the claims, definitions, qualifications, counterarguments, and relationships that must survive. This is a map of the meaning, not an outline of the sections. The revised article is allowed to change its language and rhythm, but it must still carry every important part of that map unless I have explicitly decided to remove one. ## 5. Use simple language without simplifying the argument I want a reader to understand a difficult idea without needing the vocabulary of the course, industry, or profession that produced it. When a technical term names a real distinction, I keep it and explain it. When it merely makes the sentence sound more important, I replace it. This does not mean every sentence should be short. A longer sentence may be the clearest way to show that a conclusion depends on a condition, that an objection limits rather than defeats a claim, or that two causes operate together. If splitting the sentence makes those relationships less visible, the writing has become easier to scan but harder to understand. ## 6. Choose the structure before generating the article Structure is not neutral. Beginning with an example rather than a definition changes what the reader treats as the problem; placing an objection late can make it look weaker than it is; ending with a rule can hide an uncertainty that the body of the article never resolved. I ask for several possible structures and compare what each one reveals, delays, or excludes. An AI model can produce that range quickly, but I choose the structure because it determines how the argument is experienced before the individual sentences have been written. ## 7. Draft in sections you can inspect Generating a complete article at once creates a large surface that is difficult to review. By the time I notice that the voice, evidence, or structure has drifted, the same mistake may have shaped every section. I prefer to work on a bounded part, meaning a section small enough to inspect carefully, and then check it before continuing. This makes it easier to see which language came from my notes, which connection an AI model proposed, and which change I accepted. It also lets me repair the method before it produces another thousand words in the wrong direction. ## 8. Separate drafting from criticism An AI model that has just chosen a structure or written a paragraph is likely to judge the result through the same assumptions that produced it. For an important piece, I give the stable draft to a separate critic with a narrow task: find unsupported claims, weakened qualifications, missing objections, generic language, and places where the prose became clearer by saying less. The critic may be a fresh pass or a different AI model. I do not treat its judgment as final either; the value comes from exposing problems and competing readings that were difficult to see from inside the draft. ## 9. Check evidence and edits before polishing Smooth prose can make a large change in meaning difficult to notice. After a revision, I want to know which claims moved, which terms were replaced, which sentences were joined or split, which evidence was added, and which uncertainties disappeared. I also keep observation, inference, and uncertainty separate. An observation is what a source or experience directly shows. An inference is the meaning I draw from it. When an AI model fills the distance between them with a plausible sentence, the article may become more complete while the evidence remains exactly as limited as it was before. ## 10. Approve one exact version at a time Using AI to write lets me test language, structure, and argument much faster. However, publication is not an extension of that speed. I read the final version, check its evidence and privacy, and decide whether I am willing to attach my name to the exact claims it contains. If the article changes after that decision, I review it again. Approval of one version does not cover a later rewrite, and confidence in a writing method does not authorise it to publish new work in my name. The collaboration may produce the draft, the critique, and several alternatives; the final choice of what I mean and what I am prepared to defend remains mine. --- Creator: Rahil Chamola Canonical source: https://rahilchamola.com/writing/co-writing-with-language-models Preferred citation: Chamola, Rahil. “How I use AI models when I write.” Curated Randomness, 2026.