What is a Claude Project?
Think of a Claude Project as giving Claude a dedicated brain for a specific client or task. You upload documents (Communication Codex, brand guidelines, website copy, buyer personas, past work samples) and write custom instructions. Every conversation you start inside that project begins with all that context already loaded.
It's like briefing a new team member on a client. Except you only have to do the briefing once, and Claude never forgets it and never has a bad day.
Without a project, Claude doesn't know who your client is, what their voice sounds like, or what topics to avoid. With a well-built project, you get output that's consistent, on-brand, and a lot closer to usable on the first pass. 🎯
How to set up a client Claude Project
-
Go to claude.ai > Projects > Create Project.
You'll need to be logged into the agency Claude Pro account, not a personal account.
-
Name it: "[Client Name] — [Service Area]"
Example: "Branch Furniture — Reddit Content" or "Healthy Blue — Social Media." Be specific enough that someone who wasn't involved in the setup can find it and know what it's for.
-
Write Project Instructions.
These are the standing instructions Claude gets at the start of every conversation in this project. Include:
- The client's industry, audience, and goals
- Specific output formats (e.g., "Always provide 3 caption options at 25, 50, 75, and 100 word lengths")
- Tone and voice direction (pull from their Communication Codex)
- What NOT to do (competitor mentions, sensitive topics, banned words)
- Platform-specific rules if applicable
-
Upload Project Knowledge files as .TXT, not .PDF.
This is the #1 mistake people make. Claude reads text files much better than PDFs. Export every doc as .TXT before uploading. Files to include:
- Communication Codex
- Client charter or T.A.R.G.E.T. document
- Brand guidelines (written, not the visual PDF)
- Buyer personas
- Website copy (make a "Website Copy for AI" doc with full page text)
- Past approved content samples
-
Test it with 3 to 5 real prompts.
Compare the output against real approved work. Does it sound like the client? Does it follow the formatting rules? If not, refine the instructions before using it for live work.
Case study: Buddy Punch
Buddy Punch is one of our best-built client Claude Projects, and it's a great example of what specific project instructions look like. Here's what their setup specifies:
Buddy Punch Claude Project Instructions (excerpt)
- Always prepare 3 captions in 25/50/75/100 word length variations
- Use "we/our" collective voice throughout
- Tone: conversational, not casual. Professional, not corporate.
- Hook structure: shock value or humor. Never questions.
- Advice-first approach: teach the audience something, don't sell at them
- Use specific data points when available
- Short sentences. Readable at a glance.
- No emojis. No hashtags.
- Platform adaptations: Facebook (longer, story-driven), LinkedIn (professional lens), TikTok (punchy, trend-aware), Instagram (visual hook first), X (short, direct)
Notice how specific these instructions are. "Conversational but not casual" tells Claude something. "Warm and friendly" tells Claude almost nothing. The more concrete your instructions, the better your output.
You can read the full Buddy Punch project instructions in the Google Drive doc linked below.
Keeping a Claude Project up to date
One of the most common issues with Claude Projects: they're outdated, missing info, or were set up and never touched again. Here's the workflow to keep your projects current.
The Codex update workflow
- After every client check-in where feedback comes in, ask yourself: "Does this feedback change how Claude should write for this client?"
- If yes, update the Communication Codex Google Doc.
- Export the updated Codex as .TXT.
- Replace the old Codex file in the Claude Project with the new one.
- Post in Slack (#ai-and-automations or the client channel): "Updated [Client] Codex in Claude Project — added [what changed]."
Quarterly: Review all project knowledge files for staleness. Remove outdated content. Add new approved samples. Delete old campaigns. Make sure the project reflects how the client sounds right now, not how they sounded a year ago.
💡 AE responsibility: If you're the Account Executive on a client, you own their Claude Project. When the Codex changes, you update the project. When new samples get approved, you upload them. When the project falls out of date, that's on you. This is part of the job now.
Writing project instructions that actually work
The difference between a good Claude Project and a mediocre one is almost always in the instructions. Here's how to write instructions that consistently produce good output.
✅ Specific: "Write in first person plural (we/our). Use a conversational tone similar to the examples in the uploaded website copy file. Keep sentences under 20 words."
⚠️ Vague (don't do this): "Write in a professional tone." Claude doesn't know what professional means to your client. Neither does anyone else without context.
A good set of project instructions answers these questions:
- What does this client sell or do?
- Who are they writing to?
- What words or phrases should never appear?
- What format should outputs take?
- What does great output look like? (Point to an example.)
- What platform or channel is this for?
Try it yourself
Pick a client you work with frequently. Open their Communication Codex (or ask your AE for access). Then:
- Find their Claude Project in Claude.ai
- Read the existing project instructions
- Identify one thing that's missing or out of date
- Write a more specific version of one vague instruction you find
If there's no project for a client you work with regularly, flag it in #ai-and-automations. That's a gap worth filling.
Common Claude Project mistakes
- Uploading PDFs instead of .TXT files. Claude reads PDFs poorly. Always export as text first.
- Never updating the project after client feedback. An outdated project produces outdated output. Update it every time the Codex changes.
- Writing vague instructions. "Professional tone" and "friendly voice" are not instructions. Be specific.
- Stuffing too many unrelated files into one project. Keep projects focused. One client or one service area per project. Too much mixed context confuses Claude.
- Not testing the project before using it for real work. Always run a few test prompts before the project goes live. Compare against approved output.