AI-assisted writing: a guide for people who work in an agency
Read Anthony Booth's position, then pick the page for the question you have today.
Where I stand
I have two main writing problems: drift and slop. And one main Agency problem; getting paid for it. And one thing I'm quite good at: writing to get cited.
AI writing works: in advertising tests AI-assisted writing performs better, even if clients expect the work in an hour. This guide has that research.
One warning as communicators we need to keep in our heads: my feed now looks like AI talking to AI in the comments.
"A funny thing is that I now get more insightful comments on my posts about obscure science fiction authors or modernist poets or Bauhaus architecture than ever before, but that is entirely due to AI powered slop accounts."
Ethan Mollick on X, 12 June 2026
"AI makes the pressure on hourly pricing pretty easy to understand. If work gets faster, hours become a less useful way to explain its value. But “pay for outputs” isn’t a complete answer either. You still have to define the output."
Ryan Bozeman, content marketer, on X, 3 October 2026
Drift is just a weak prompt
When there is nothing to anchor the LLM it slightly misses the point and adds words but no value. Mike's prompts overcome that issue. Jano is right, too - content recycling will kill us - unique facts or insight will help our client's content stand out and rank.
"AI rewards structure. Bad thinking plus AI just gets you bad answers faster."
Samson Marks, UK agency founder, on X, 30 September 2026
Slop is just lazy 'send'
It takes James just 7.5 seconds to recognise my sloppy writing and ask the difficult question 'Did you actually read it??' What makes him mad is that (feels like) it took him more effort to read it than it took me to write it. My editing, judgment and fact checking should be at least 20% of the time spent on the task.
"If you use AI to write me a long email asking me to help you 1) I can tell 2) I’m never going to help 3) I’m going to block your email Have some respect for people"
Tom Goodwin, advertising author, on X, 27 August 2026
"AI will be especially fatal to writing that takes too long to get to the point. Now there will always be a more concise alternative, and 99% of readers will choose it."
Paul Graham on X, 5 September 2026
AI tells have changed: Don't worry so much about Em Dashes (retired by Chat GPT), but about sentence structure like; "It's not X, but Y."
""it's not x it's y" used to be what we said to emotionally wrap up an argument, double down on a serious point. Using it every third sentence feels sterile."
Alexander Sheppert, physician and AI engineer, on X, 5 October 2026
"I was thinking about how I recognize AI writing, and one big tell is excessively colorful verbs. A handful of journalists might write that a proposal "drew" 100 votes, but any normal person will just say "got"."
Paul Graham on X, 19 August 2026
"Its fitting that one of the tells of AI writing is over-assigning agency and action to inanimate objects: "the code knows it now," "the plan remembers," etc."
Ethan Mollick on X, 17 September 2026
Stop using a chat bot. I use these reusable skills in Claude co-work. The most important is the 'declaudification skill'.
Don't be a meat puppet. Don't outsource your thinking. My best work comes through a 2-way process with the LLM and my hook and thinking evolves. I am trying to make Claude (better, to use a 2nd model) a harsher critic than James. Often what you think changes while you are deciding what to write.
"Each of them has done an absolutely incredible job. [...] But at the same time not a single slide is good enough."
Tom Goodwin on X, on a presentation that he gave to three models to design, 30 September 2026
Ask for options: Generative AI picks the most plausible words. So force it to choose different ones. Then mix and match.
"AI can write you 50 headlines in ten seconds. Someone still has to pick one and delete the other 49. You are the tastemaker."
Copyhackers, copywriting school, on X, 6 October 2026
Don't start writing until the research is done. For Fundrecs I found EVERY single post, whitepaper, webpage and comparison before asking the LLM to analyse or write anything. It was over 2,500 documents.
Employ writing skills and create your own voice
- structure-first: who is it for, why, what is the one thing they will leave with
- write-to-anthony: caveman/terse/bullets - in chat with LLM
- write-as-anthony: trained on my voice
- write-house-copy: ads/copywriting - trained on Ogilvy
- write-client-docs: developed with James for Board-level (English, Consultant)
- simple: ASD - for internal/project [new]
- remove-ai-tells
Try these things
- train your writing agent on a library of great writing you love. (mine is called William Shakespeare)
- use a list of the 'grammatical' constructions that AI uses most frequently and set a -ve match list
- set gates on output to force it to read itself
- force it to declare its context before writing (did you read x and y)
- edit for length and repetition and factual density and remove self-narrative from the headlines
- get it to list the issues in its own writing
- have a final human editing pass - I use an artifact and edit live in side by side windows
- ask it to do 'diff' versions (get it to tell you what it changed)
- force it to read it as its intended subject and score its impact, clarity and humanness
- force it to check and publish its sources
- talk to the LLM (I have about 100 voice notes that I have used as input to documents)
- over several drafts you lose facts, so don't regenerate copy; version and edit. (That was a painful discovery at 3am James....)
Get your LLM to talk to you more clearly: Karpathy wants his LLM to write in ASD-STE100 - which is a simple writing style. Here is an example used in a market report for a Health ingredient client.
Write to get cited
Organic rankings don't deliver as much traffic any more. But getting ranked will deliver citations. Here's how I do that. I use Google's guide to design the brief, research structure, create unique data, FAQs. Here's an example of a Blog for a UAE client where I score my page versus the existing citations on Google's own score.
(This guide assumes you use a co-work style AI, with skills, connectors and memories. Rather than the chat-bot.)
Anthony Booth, Articulate, 7 October 2026
Three of these points, checked against the research. In advertising tests, AI copy matched staff writers on sales in three email trials, and a model tuned on past test results beat human experts on subject-line clicks by 32%. A bare prompt did no better than the experts. A human edit of AI emails added no measurable sales in the one trial that tested it. Karpathy's post is about asking a model to explain its answer in ASD-STE100 so that he can read it faster. The detail is on What works and What is new. The full check of all twelve practices is on How this guide was made.
Pick your question
| If this is your question | Read this |
|---|---|
| A client wants a saving from AI. What do I say? | Why AI, which opens with a script |
| I am a writer. What keeps me employed? | Day to day, the copywriter block |
| Is this draft good enough to send? | Levels of quality, with one piece shown at each level |
| What can go wrong, and who pays? | The issues |
| What do I do differently on Monday? | Day to day |
All ten pages
| Page | The question it answers |
|---|---|
| 1. Why AI | Why is AI now the normal case for agency writing, and what saving is real? |
| 2. The issues | What goes wrong, and who has the record of it? |
| 3. What people say | What do writers, clients and readers say in their own words? |
| 4. Levels of quality | How do you tell slop from work that sells? |
| 5. Tools now | Which tools do agencies use, for what, at what price? |
| 6. What is new | What changed between April and October 2026? |
| 7. What has changed | Which advice from 2023 is now false? |
| 8. What works | Which practices have a measured result? |
| 9. Day to day | What do you do on Monday, in your role? |
| 10. How this guide was made | Who wrote it, how was it checked, and what is weak? |