How I Set Up Claude Projects So Client Work Stops Leaking Context
One project per client, a style file, a facts file, and a hard rule about what never goes in. This is the setup that made AI assistance safe for agency work.
The moment AI assistants became useful for client work was the moment context stopped living in prompts. My setup is deliberately boring: one Claude Project per client, two permanent documents, and one rule enforced without exception.
Document one is the voice and style guide, tone, banned phrases, reading level, and three past paragraphs the client loved. Document two is the facts file: product details, pricing, dates, and the correct spelling of everyone's name. Every embarrassing AI hallucination I have shipped traced back to making the model infer a fact that should have been written down.
The hard rule: drafts, feedback, and anything covered by an NDA go into the project only if the client has agreed in writing to AI processing. For the two clients who have not, I run a separate project with fake names and no attachments, which is less convenient and exactly as safe as the contract requires.
The payoff shows up in week three, when you ask for a revision and the model already knows the client hates the word solutions. My editing time on routine deliverables dropped by roughly a third, and the tone is consistent across whoever on the team writes that day.
One refinement worth stealing: keep a running corrections note in the project and append every fix you make by hand. Once a month, fold the corrections into the style file. The project slowly learns your client the way a good account manager would, and unlike a good account manager, it never gets poached.