Building your first real Project: a worked example
Different jobs, one repeatable setup sequence
After this, you'll be able to build a full Project (role Instructions, Knowledge Files, and a clear name) for your own recurring work by following one repeatable sequence.
Before you start
Complete Memory vs Projects: why you need both first; this lesson builds on that routing rule by assembling Instructions and Knowledge Files into one complete Project you set up end to end.
The idea
Every Project, for writing or client work or research, follows the same three-step sequence: role Instructions, then Knowledge Files, then a clear name. You have learned the parts; this lesson assembles them into one repeatable move. The content changes per job, but the order never does.
Walk through a weekly newsletter. The Instructions hand Claude a role ("You are my newsletter co-writer for busy parents; warm tone, one-line hook, one takeaway, no jargon"). The Knowledge Files are your three best past issues, and the name is "Newsletter, Weekly" so it sorts predictably.
Here is the before and after: Before, someone names a Project "stuff," writes one vague Instruction ("help me write"), uploads nothing, then wonders why the output is generic and they cannot find it a week later. After, the three-step sequence produces a draft that needs light edits, and the Project is one click to find.
Now try it: pick one kind of work you do weekly and run the three steps in order, a role-based Instruction, two reusable Knowledge Files, then a name like "Client, Acme" or "Newsletter, Weekly".
The content changes per job; the three-step sequence is what you reuse to build any Project fast.
Try it (14 min)
Watch out for
- Skipping the role line in Instructions; 'help me write' is weaker than 'You are my newsletter co-writer for busy parents'
- Uploading no Knowledge Files and expecting your voice; Claude needs examples or a style doc to match how you sound
- Naming a Project 'stuff' or 'test'; a vague name costs you every time you scan for the right Project later
- Running the steps out of order or skipping one; the sequence is Instructions, then Knowledge Files, then name, every time
- Building the Project and never sending a test chat; verify the setup works before you rely on it for real work
Paste this into Claude
Help me build a complete Project from scratch using the three-step sequence: role Instructions, then Knowledge Files, then a clear name. Here is the recurring work this Project is for: [describe it, e.g. "writing my weekly newsletter for busy parents" or "preparing briefs for my consulting clients"]. Please produce: 1. INSTRUCTIONS: a role-based set of Instructions under 120 words, written as direct commands, that gives you a clear role, the audience, the tone, and the output format I should always get 2. KNOWLEDGE FILES: a short list of 2 to 4 documents I should upload as background for this work, and one sentence each on why 3. NAME: three naming options that follow a consistent convention so this Project sorts cleanly next to others I will create Then give me a one-paragraph "first chat" message I can send inside the Project to test that the whole setup works.
What good looks like
- Claude produced role-based Instructions under 120 words written as commands
- Claude listed 2 to 4 specific, reusable Knowledge Files with a reason for each
- The naming options follow a consistent, scannable convention
- You actually created the Project, pasted the Instructions, and uploaded at least one file
- Your first test chat inside the Project produced output that matches the role, audience, and format
When this breaks
- Breaks when you skip the Knowledge Files step for voice-sensitive work, because Instructions alone describe the tone but examples are what let Claude actually match it.
- Breaks when naming is inconsistent across Projects, because once you have several, you cannot scan and find the right one, so you lose the speed the setup was meant to give you.
AI can help with this
Paste this into Claude: 'Help me build a Project for [your recurring work]. Give me (1) role-based Instructions under 120 words as commands, (2) a list of 2 to 4 Knowledge Files to upload with reasons, and (3) three consistent naming options. Then write a first test message I can send inside the Project.'
You can now
You can complete the lesson outcome in a real Claude chat, Project, Artifact, Connector, Desktop, or Code surface.
- You can verify that claude produced role-based Instructions under 120 words written as commands.
- You can verify that claude listed 2 to 4 specific, reusable Knowledge Files with a reason for each.
- You can verify that the naming options follow a consistent, scannable convention.
- You actually created the Project, pasted the Instructions, and uploaded at least one file.
Key takeaways
Every Project follows the same three-step sequence: role Instructions, Knowledge Files, then a clear name. The content changes per job; the sequence is what you reuse to build any Project fast.
- Follow one sequence for every Project: role Instructions first, then Knowledge Files, then a clear name.
- Give Claude a role in the Instructions, not a vague 'help me' request.
- Add Knowledge Files (past examples, a style doc) so Claude matches your voice, not just your description of it.
- Use a consistent naming convention so Projects sort and scan cleanly once you have several.
- Run a test chat after setup to confirm the output matches the role, audience, and format before you rely on it.