Running Claude Code (or Hermes/GPT) for client work: one workspace, or one per client?

Question for anyone who has AI wired into how their agency actually runs.

I built a workspace that holds my business context: positioning, offer, ICP, voice, strategy, numbers. It works because I stop re-explaining my business every session. Everything I ask gets answered through that lens.

Now I'm bringing client work into it and I've hit a structural problem.

If clients live in the same workspace, my own rules bleed into their work. My positioning rules exist to protect my sales process. Applied to a client's copy they are simply wrong, and the output still reads confident, so I don't catch it.

If each client gets a separate workspace, that's solved. But then I lose the thing that makes an agency worth more each year: the pattern library. Which angles work. Which structures convert. Separate workspaces means separate brains that never learn from each other.

So it's contamination on one side, amnesia on the other.

Where I'm leaning: one workspace, each client isolated in their own folder, plus a shared library where proven patterns get promoted up with client-specific details stripped out.

I'd like to hear from people who are already running this.

One workspace, or one per client?
How do you stop your own voice and rules from leaking into client output?
Does anything in your setup learn across clients, or does each client start cold?

Even a one-line answer helps. Happy to report back on what I land on.

5 comments

Running Claude Code (or Hermes/GPT) for client work: one workspace, or one per client?