The Right Way to Write With AI: Build a Personal Language Model

The right way to write with AI is to separate thinking from repetition. A person creates the experiences, opinions, judgments, stories, and new connections. AI retrieves approved material, reshapes it for a format, and repeats it consistently. Without that source library, the model can only produce a polished version of what everyone already knows.

The Method in One Minute

1. Choose a small number of topics you are willing to think about for years.
2. Write down real experiences, decisions, failures, observations, and opinions in your own words.
3. Separate shared knowledge from personal evidence and original ideas.
4. Store approved passages with topics, dates, sources, permissions, and context.
5. Ask AI to retrieve and adapt only relevant approved material for a defined audience and channel.
6. Review every personal claim, quotation, number, and implication before publishing.
7. Return the approved result, audience response, and corrections to the library.

Nicolas Cole calls the accumulated library a personal language model or digital brain. This is a content system, not a replacement foundation model. Its value comes from the specificity and provenance of the material inside it.

Watch the Full Conversation

Source and credit: The Right Way To Write With AI, published by Greg Isenberg on 21 September 2026, featuring writer and Typeshare co-founder Nicolas Cole. Cole also published the framework in his own essay, which is used below as the primary written source.

The Three Tiers: Commodity, Personality, and Original

Commodity content is not automatically bad. A beginner still needs the definition, checklist, or first principle. The problem appears when a content strategy stops there. AI can produce unlimited competent explanations of conventional wisdom, so generic coverage has little durable differentiation.

Personality content makes the shared idea attributable. Instead of writing, “Founders should speak to customers,” record the specific interview that changed your roadmap, what you misunderstood, and which sentence made you reconsider the product. The reader receives both the common lesson and evidence that a real person earned it.

Original content is rarer. It emerges from sustained attention to a narrow area, not from adding unusual adjectives to a familiar claim. A useful originality test is simple: what does this piece add that a careful reader could not recover from the standard sources?

Ownership Is Association, Not Mere Authorship

Cole argues that much of language is shared. No writer owns ordinary words or every familiar combination of them. In his framework, practical ownership grows when an audience repeatedly associates a specific idea, story, phrase, or method with a person.

That makes volume useful, but only when the repeated material carries a recognizable point of view. Publishing the same commodity advice at high frequency builds reach, not necessarily ownership. A stronger sequence is:

1. Choose the idea you want to explore.
2. Attach real evidence and a personal interpretation.
3. Repeat the core position across different examples and formats.
4. Update it when new evidence changes your mind.

Association is not proof of truth. A writer still needs sources, correction history, and clear boundaries between experience, inference, and verified fact.

Build the Digital Brain as a Source Library

The digital brain should be more structured than a folder of finished posts. Store small source records that remain useful across formats:

A transcript is raw material, not automatically approved language. Extract candidate passages, check them with the speaker, preserve the original recording, and record any edits that change meaning.

A Human and AI Writing Workflow

Human responsibilities

• Choose the question worth answering.
• Supply first-person evidence and make the judgment.
• Decide what is confidential, uncertain, or not ready to publish.
• Create the new framework or connection.
• Approve the final claims and language.

AI responsibilities

• Retrieve relevant approved passages and their sources.
• Find repeated ideas, gaps, contradictions, and unused examples.
• Build outlines for a specific reader and outcome.
• Adapt one approved source into a newsletter, post, script, FAQ, or sales aid.
• Flag unsupported personal claims instead of filling the gap.

A useful draft request is: “Use only the attached approved records. Cite each personal claim to its source record. Do not create anecdotes, opinions, results, or quotations. Mark missing evidence with [SOURCE NEEDED]. Preserve the central argument while adapting it for this audience.”

Define Voice With Three Observable Dials

Cole reduces voice to three controllable variables:

1. Word choice: plain or technical, formal or conversational, literal or metaphorical.
2. Sentence length: short and forceful, long and cumulative, or a deliberate mixture.
3. Sentence structure: direct claims, questions, fragments, stories, contrasts, lists, and transitions.

Turn each variable into examples from approved writing. “Sound confident” is vague. “Lead with a direct claim, use short paragraphs, prefer concrete verbs, avoid inflated adjectives, and close with a practical decision” is testable.

Voice is still a minority part of differentiation. Perfectly copying rhythm cannot compensate for an empty idea. Build the evidence and point of view first, then teach the machine how to preserve their expression.

Choose the Time Horizon Before You Write

Cole places content on a spectrum from timely to timeless. Both can be valuable, but they should be judged differently.

Do not judge a timely post by whether it attracts readers three years later, or judge a foundational essay by its first 24 hours. Record the intended horizon in the brief so the format, research effort, and measurement match the job.

From Personal Language Model to Company Brain

The same method can become company infrastructure. A strong company brain contains founder thinking, customer language, research, product decisions, case evidence, approved terminology, objections, and corrections. Teams can then use AI to retrieve the right source for a campaign without asking the founder to repeat the same explanation every week.

The moat is not the vector database or the chatbot attached to it. Competitors can buy similar software. The harder asset to copy is the accumulated judgment: why the company made a decision, what it learned from customers, which claims survived scrutiny, and how its point of view changed.

A 30-Day Implementation Plan

1. Week 1, collect: choose three topics and gather 20 strong pieces, transcripts, notes, or voice memos you own.
2. Week 2, structure: extract 50 source records, label the content tier, attach evidence, and remove anything you cannot publish.
3. Week 3, articulate: write five new personality pieces yourself. Focus on decisions, mistakes, details, and changed beliefs.
4. Week 4, scale: let AI adapt one approved piece into three formats. Compare each draft with the source, correct it, and save only the approved result.

After 30 days, review which records were repeatedly useful, which prompts created unsupported claims, and which ideas deserve deeper original work. The goal is not maximum content volume. It is a better source library and more time for the next thought.

Video Chapters

Sources and Links

• Greg Isenberg and Nicolas Cole: The Right Way To Write With AI
• Nicolas Cole: The Right Way To Write With AI (primary written framework)
• Typeshare and Nicolas Cole on Typeshare
• Nicolas Cole's official website
• Ship 30 for 30 (episode link)
• Brex Financial OS (episode sponsor link)