The draft is shared. The point of view is yours.
The model can carry the draft. The argument stays mine.
The friction was never creation. It was output. I have never been short of ideas. I have ADHD, and the gap between having a thought and publishing it is where most of mine died. Not for lack of quality. For lack of throughput. So the choice in front of me was never AI or authentic. It was AI or silence. I chose output. Most of what is on this site would still be a voice note on my phone.
Then removing the friction created a different problem. When you can produce at speed you lose the excuse that you will get to it eventually, and the bottleneck moves from execution to taste. Taste is the harder one. So everything below is friction I put back on purpose: the parts I refuse to hand over, and the checks that stop a fluent draft from quietly becoming someone else's argument.
Give the collaboration a spine
Write the thesis before opening the chat.
Start with one sentence you genuinely believe. The model may challenge it, sharpen it, or generate alternatives. It should not manufacture your conviction after producing a page of plausible prose.
The thesis does not have to be perfect. It has to be yours. If the work changes your mind, write the new thesis explicitly instead of letting it drift sentence by sentence.
Build a perspective packet, not a biography dump.
Give the model the smallest relevant bundle of audience, evidence, examples, tensions, taste references, constraints, and privacy limits.
Outcome: What should this piece let the reader understand or do?
Audience: Who is it for, and what do they already know?
Thesis: What do I currently believe?
Evidence: Which sources, experiences, or artifacts may support it?
Tension: What is the strongest objection or unresolved question?
Taste: Which qualities should the writing preserve or avoid?
Boundaries: What is private, uncertain, forbidden, or out of scope?
Output: What shape should the next artifact take?
Failure policy: What should happen when evidence is missing?Authored words are law until explicitly reopened.
Use my exact words where they already carry the meaning. Thesis lines, personal experiences, verified claims, and phrases I deliberately wrote should not be silently replaced. The model may flag a conflict and offer alternatives.
I am not protecting every comma. I am protecting the reason the piece exists from being smoothed away.
Keep selection visible
Ask for choices before asking for prose.
Structure is already an argument. Request distinct outlines, openings, metaphors, or counterarguments before letting one generated path harden into the piece.
Selection is authorship. A model can create breadth cheaply; that makes the choice between options more important, not less.
Draft in stages.
Move through thesis, structure, section drafts, evidence review, critique, and rewrite. Keep one bounded artifact and one active writing job at a time.
- Human thesis
- Scoped perspective packet
- Model-generated options
- Human selection
- Bounded section draft
- Evidence pass
- Independent critique
- Human final edit
Separate the writer from the critic.
The critique should inspect a stable draft against explicit criteria. For consequential work, use fresh context or an independent model rather than letting the draft approve itself.
Disagreement is useful evidence. Compare the claim, source, edit, and uncertainty. Do not settle a conflict because one model sounds more certain.
Facts are not style material.
Never allow invented numbers, quotations, experiences, or citations to make the paragraph flow. Unsupported gaps stay visible until they are verified or removed.
I ask the model to separate what is observed, inferred, and unknown. That one distinction prevents a surprising amount of confident fiction.
Take the work back
Restore human rhythm.
Read aloud. Vary sentence length. Replace generic abstractions with concrete nouns and verbs. Remove anything that sounds competent but carries no lived observation.
The goal is not to trick an AI detector. The goal is to sound like a particular person who noticed a particular thing. Mechanical “humanisation” tricks create another kind of generic prose.
Inspect the diff, not just the polish.
Ask which claims changed, which authored lines moved, what evidence appeared, and what remains uncertain. Smooth prose can conceal a semantic rewrite.
This matters most when the draft sounds better. Surface improvement makes it easier to miss that the argument became safer, broader, or less true.
Own the final sentence.
Verify citations, privacy, implications, and disclosure needs. Shared labour does not transfer responsibility for publication.
A complete starting packet
Thesis:
The model is not the product. It is becoming infrastructure.
Audience:
Product builders who confuse model capability with product experience.
Evidence:
Three observed product failures and one credible counterexample.
Voice:
First-person, discovery-first, punch / explain / punch / ground.
No expert-at-a-podium posture.
Locked lines:
The model is not the product.
Speed is the tool. Judgment is the craft.
Next output:
Three distinct four-part outlines. Include the strongest counterargument.
Do not draft prose yet.
Failure policy:
Mark unsupported claims [SOURCE NEEDED].
Do not invent examples, numbers, or quotations.Signs the model has taken over
Stop and recover the point of view if…
- The thesis appeared only after the prose.
- Every sentence is equally smooth.
- The evidence could support almost any argument.
- No authored line survived unchanged.
- The same pass drafted and approved the work.
- You cannot explain why a section exists.
- You would not defend a claim without mentioning the model.
Writing has always involved editors, conversations, references, and tools. What matters is knowing where the judgment lives.
Let the draft be shared. Keep the point of view yours.