Skip to content
Multigrid

Writing, media and creative work

Where a language model genuinely helps a writer, where it flattens the work, and the specific edits and checks that tell the two apart.

Writing is the task these models were built for, which is exactly why the advice about it is so bad. Most of what is published amounts to “use AI to write faster” plus a list of prompts, and it skips the only question a working writer has: what does the machine actually contribute, and what does it quietly take away?

The short answer, which the rest of this cluster spends twenty pages earning: a model is very good at operations on text you already have and unreliable at operations that require judgement about what is worth saying. It will restructure, compress, translate, extract, transcribe and diagnose. It will not select. Selection — deciding what to cut, what to claim, and what is not worth publishing at all — is the part that was always the work, and it is the part that did not get cheaper.

These pages are concrete about the difference. There are eleven named tells of unedited machine prose with the edit that removes each, a style guide rewritten as constraints a script can check, a research workflow in which the model is never permitted to supply a fact, and a definition of slop that does not rest on who or what typed it. Where a claim needs arithmetic, the arithmetic is on the page.

Where AI Fits in a Real Writing Process

Research, drafting and editing are three different jobs. A model is excellent at one of them, dangerous at another, and mixed at the third.

8 min read

Editing AI Prose Into Something Publishable

Eleven specific tells of unedited machine prose, each with a real before and after, plus the order to edit in and what editing cannot fix.

14 min read

Research With AI Without Inheriting Its Mistakes

A workflow in which the model never supplies a fact: you find the sources, it extracts from them, and every extraction is verified as a literal substring.

11 min read

Outlining and Structuring Long Pieces

Build the argument as bare claims first and let the model attack it, then generate prose one section at a time against a claim you wrote.

10 min read

Headlines, Titles and Subject Lines

Generate wide, then select against your own archive — with the arithmetic showing why most publishers cannot A/B test their way to an answer.

10 min read

A Style Guide a Model Can Follow

Rewrite house rules as constraints something can check, ship a checker script, and keep the judgement rules where they belong — with a human.

11 min read

Telling Readers That AI Was Involved

The four positions publishers actually take, the threshold that decides which one applies to a piece, and a disclosure template to adapt.

9 min read

Plagiarism, Paraphrase, and Where AI Output Sits

Three different questions get asked in one sentence — academic, copyright and editorial. They have different answers, and mixing them is why the argument goes nowhere.

11 min read

Dialogue and Subtext: What Models Are Bad At

Five specific dialogue failures, each shown on the page as generated-style text and a repair, plus the single mechanism that produces all five.

11 min read

Fiction With AI: Tools, Limits and the Craft Question

What a next-token model actually contributes to a manuscript, which novelistic tasks that property ruins, which it leaves untouched, and the craft objection stated at full strength.

11 min read

Writing an Illustration Brief for an Image Model

The six slots every image brief needs, the art-direction vocabulary that changes an output, and the one-variable iteration loop that finds out which word did the work.

10 min read

Video Scripts, Storyboards and AI-Assisted Production

Stage by stage through a video pipeline: where machine assistance removes real hours, and where it adds a review cost bigger than the time it saved.

11 min read

Podcast Production With AI

Transcription, show notes and chapters, each with a stated quality bar — including the arithmetic showing what a 5% word error rate costs you on one episode.

11 min read

Localising Creative Work, Not Just Translating It

Idiom, humour, register and cultural reference are the four things machine translation loses — and the automatic metrics are built to penalise the fix.

10 min read

What Image Platform Terms Let You Do With Output

The six questions that decide what you may actually do with a generated image, where each is answered in a typical terms document, and what to record.

10 min read

Building a Personal Voice Corpus

Select passages from your own archive, use them as few-shot exemplars, and test the result with a stylometric measure that does not depend on your opinion of it.

12 min read

An Editorial Workflow With an AI Step and a Human Gate

One gate, held by a named person, with five checks that run on every piece — and the arithmetic showing what the gate costs and when it pays for itself.

11 min read

Writing Documentation With AI Assistance

Extract the API surface from the source first, write against the extraction, then fail the build on any identifier in the prose that the extraction does not contain.

12 min read

Cadence, Volume and the Slop Problem

Sending more often multiplies churn linearly while adding reach that saturates — an argument made with arithmetic on the page rather than with anybody's dashboard.

11 min read

What "AI Slop" Means, and How Not to Make It

Slop is a set of five properties, not a question of who typed it. Each property has a specific fix, and the test applies to human work just as well.

12 min read

Other topics