AI-assisted writing: a guide for people who work in an agency

Where I stand

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.

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.

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."

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.

Ask for options: Generative AI picks the most plausible words. So force it to choose different ones. Then mix and match.

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

Pick your question

All ten pages