Louise Brogan had a Claude Project open on her screen, and she read out one line of its instructions that stuck with me: quotes are data.
If you write content for clients, you know the problem that line solves. You hand an AI tool the topic, the client’s industry, and a few notes from your kickoff call, and you get back a polished LinkedIn post that sounds fine and belongs to nobody. Then you spend the next half hour rewriting it so the client recognizes themselves. Or worse, you don’t catch the anecdote the model made up, and the client does.
I’ll be honest with you. When Louise joined me on The AI Social Playbook, I told her this was something I’d been doing hit or miss. I trained an assistant on my own writing years ago – I had over a thousand articles to feed it, so that part was easy – and that solved how I sound. It never solved the problem Louise solves every week, which is staying inside what one specific client actually said. She manages that with instructions rather than better prompts.
What Louise built, and why
Louise runs a LinkedIn content business for professional-services clients: leadership coaches, learning and development firms, financial services, and a law firm. Those clients are wary of AI content to begin with, and she pointed out that LinkedIn doesn’t want AI-generated posts either. So she uses Claude as a tool, not as the writer.
Every client gets their own Claude Project. Each client records a 30-minute video with her every four to six weeks, and the transcript of that session goes into the project. From there, Claude suggests post ideas, a human copywriter on her team writes the actual posts, and the client approves everything before it’s scheduled.
The part worth copying is how she wrote the project instructions. You can watch her walk through the whole thing in Episode 14 of The AI Social Playbook, but here is the mechanism, step by step.
Step 1: Give one project one client
Create a separate Claude Project for each client, named for that client. That sounds like housekeeping, and it’s actually your first guardrail. A project only knows the instructions, files, and transcripts you put in it, so one client’s phrasing never leaks into another client’s posts.
Step 2: Write the instructions as rules, not requests
Louise’s instructions start with the job: who the client is, what they do, and what the project produces for them. In her case, that’s LinkedIn content plus YouTube content for this particular client. Then come the rules, and most of them tell Claude what not to do.
Set the locale. Louise is in the UK, and this client is American, so the instructions say American spelling. It’s a tiny line, and it’s the kind of mistake a client spots in the first sentence.
Name whose voice this is. The instructions tell Claude this is the voice of the client.
Never invent client stories. This closes the most dangerous gap. When a post needs an example, and the transcript doesn’t have one, a model will happily supply something plausible. Your client then reads a story about their own business that never happened.
Quotes are data. My read of this rule: a line from the transcript is source material to preserve exactly as spoken, not wording to improve. Clean up the grammar, and you’ve changed what your client said.
Leave a placeholder. Where information is missing, Claude marks the gap so the team can check, instead of filling it. That gives the model a sanctioned way to say it doesn’t know.
Louise adds one more control above all of this. Her Claude is set up so it has to check with her before it does anything, in a project or in a regular chat. She called it the control nature of the entrepreneur. I’d call it the cheapest insurance policy in the whole system.
Step 3: Back the rules with knowledge files
The instructions point Claude to three knowledge files in the project. One is a client overview. One maps the content pipeline: the client records, raw files go to the editor, transcripts come into the project, the team creates the written assets, the client approves. The third covers brand style and content principles, down to rules like graphics never covering the client’s face.
The pipeline file does quiet but important work. Because Claude knows its place in the process, it doesn’t try to do the whole job. When Louise asked for content in the demo, Claude asked clarifying questions and offered only the outputs her instructions allow. It didn’t suggest TikTok videos because she’s never asked it for TikTok videos.
Step 4: Ask for ideas, not posts
This is the rule that surprised me most from someone this deep in the tools. Louise asks Claude to come up with ideas for the posts rather than writing the posts. A person writes them. That keeps the model doing what it’s good at, finding the usable moments in a 23-minute conversation, and keeps the sentences that carry a client’s name in human hands.
Where it breaks
A few limits worth naming before you build this. Louise was running Claude on Opus 5.5, with the Descript and Canva connectors doing the heavy lifting on video and graphics, so results on a different setup will vary. She showed the key lines of her instructions on screen, but not the full text, so anything beyond those lines you’ll have to write yourself.
She was also candid that parts of her process wouldn’t survive growth. She edits the client videos herself because she enjoys it, and said that wouldn’t be possible at 20 clients. Her Descript workspace is, in her words, a little less organized than the rest, and next on her to-do list.
And none of these rules replace a read-through. A model can still paraphrase a quote while believing it’s preserved it. The rules shrink the problem; your copywriter and your client’s sign-off close it.
The rule underneath the rules
We spend a lot of energy teaching AI what we want. Louise spent hers telling it what it isn’t allowed to do. That’s the part that transfers to whatever tool you’re using next year: the more your content depends on someone else’s words, the more your instructions should read like a set of boundaries.
Start with one client. Write the rules above into a project, add the three knowledge files, and run one transcript through it this week. Then compare the post ideas you get against the transcript, line by line, and see whether anything showed up that your client never said.