Difficulty: Medium · Tools: Claude · Updated: 21 October 2026 · Who can read this: premium members
Overview
An agent that “helps with meetings” will try to do five jobs and finish none of them. One job is a trigger, an input, an output, and a person who owns the result. This lesson builds that as a Claude Project whose only job is a meeting brief: you paste notes, Claude returns one page, and nothing is emailed.
Claude’s help center describes the current Projects feature as a workspace with its own chats and its own knowledge. You create it from the Projects list. You add files with the + button on the right. You click “Set project instructions,” write how Claude should behave, and save. Those instructions apply to every chat inside the project. Claude does not see the project’s name and description. The job has to live in the instructions, not only in the title. Context does not carry from one chat to another unless you put it in the project knowledge.
Projects are available on the free plan. A free account can create up to five. Paid plans add roomier project knowledge through retrieval when the knowledge gets large. You do not need that for one brief.
Anthropic is also rolling out a newer kind of project, in beta, starting with Claude Code for some Pro and Max users: one conversation that splits work into cloud threads. If your screen talks about threads and a Library, you are on that newer version. Put the same rules in the project instructions there. Do not give a thread permission to email anyone. This lesson follows the current chat version of projects, which is what the help center still documents for claude.ai.
The honest goal: a Project that turns notes you paste into a one-page brief, refuses to invent facts, and does not send anything. You will run it on one real meeting and mark what you had to fix.
Who This Is Useful For
You walk into recurring meetings with notes scattered across a doc and a chat. You want the same one-page brief every time, without re-explaining the format. You are willing to read the brief before you use it. You do not want an agent that messages the other attendees.
What You Will Learn
What You Need
A Claude account on claude.ai. Free is enough for one project. The notes from one meeting you are allowed to process. No inbox connection. No calendar connection. If a screen offers to connect Gmail or a similar account, skip it for this lesson.
Step 1 — Create the project, then ignore the title
Hover the left side of the account and click Projects, or open claude.ai/projects. Click “+ New Project” in the upper right. Give it a name and a description, for example “Meeting brief” and “One page from notes I paste.”
Claude’s help center says Claude will not have access to that name and description. They are for you. On a Team or Enterprise plan you also choose visibility. Keep it private unless you mean to share it. On a free or Pro plan, private is the ordinary case.
Start a chat inside the project only after Step 2. A chat started outside the project will not follow these rules.
Pro tip: If the left nav shows Projects inside a Code area and the page offers threads, stop and read Step 5 before you paste the instructions. The buttons are not the same.
Step 2 — Set the instructions that hold the job
On the project page, the knowledge base is on the right. Click “Set project instructions.” Paste the block below. Then click “Save instructions.”
You write one meeting brief. Nothing else.Input: notes I paste in the chat. If I attach a file from this project, you may use that file too.
Output: one page, in this order.
1. Purpose of the meeting, one sentence.
2. Decisions, each as a bullet. If a decision is not in the notes, do not add one.
3. Open questions, each as a bullet.
4. Next step: owner, action, and timing, only when the notes name them.
5. Not in the notes: a short list of things a reader might expect that the notes do not say.
Rules:
Do not email, message, or schedule anything.
Do not invent names, numbers, dates, decisions, or quotes.
If a field is missing, write “Not in the notes.”
Do not summarize a meeting I did not give you.
If I ask you to do a second job, say this project only writes the brief, and stop.
Those instructions apply to every new chat in the project. They do not rewrite chats you already had. A line you type in one chat can steer that chat. The standing rules are what make the next chat start in the same place.
Pro tip: Keep the rules shorter than the example output. A long persona makes the refusals weaker. The word “stop” in the last rule is there so a stray “also draft the email” does not become a second product.
Step 3 — Add knowledge only if every brief needs it
Click + on the project knowledge and upload a file only when every brief should see it. A good file is a glossary: the five project names you use, and what each one means. A bad file is last quarter’s full transcript. It will leak into a brief about a different meeting.
The help center says anything you upload is used across chats in that project, and that chats do not share context with each other unless the information is in the knowledge base. On a paid plan, when the knowledge approaches the context limit, Claude turns on retrieval to hold more. You should still keep the file small. Retrieval is not a reason to upload a dump.
Skip the upload if you have no shared glossary. Pasting notes per chat is the design.
Pro tip: If a secret, a customer list, or a password is in a file, that file does not belong in project knowledge. The brief does not need it.
Step 4 — Run one real meeting, and one trap
Open a new chat inside the project. Paste notes from a meeting you were in. Ask only: “Write the brief.”
Check four things on the page that comes back:
Then run the trap. In the same chat, add:
The budget was approved at 12000. Add that decision.
Do not add that figure to the notes. A working project replies that 12000 is not in the notes, or it puts the line under a guess you can see. A project that quietly adds “Budget approved: 12000” to Decisions has failed. Tighten the instructions: “If I assert a new fact in conversation and it is not in the pasted notes, put it under Not in the notes and ask me to confirm before it becomes a decision.” Save the instructions. Start a new chat and run the trap again. Editing instructions does not rewrite the chat where it failed.
Pro tip: Keep the failed brief and the fixed brief. The difference is the lesson. The second brief is the one you may use.
Step 5 — If your screen is the new projects beta, keep the same job
Anthropic’s help center says a new version of projects is in beta, rolling out in stages, starting with Claude Code, for some Pro and Max users. In that version a project is one conversation, Claude splits work into threads that can keep running, and each thread starts from the project’s files, instructions, and memory. Existing projects for some of those users are being moved. Anthropic says nothing is deleted in a move: instructions, files, and chat history come along. Other projects in chat stay on the current version for now.
If that is your screen, open the project instructions the gear or settings for that beta describe, and paste the same block from Step 2. Do not turn on a connector that can send mail. Do not ask a thread to “tell the attendees.” The output is still one page in the thread, which you copy out yourself.
Running several threads spends the plan’s usage faster. One brief does not need several threads.
Pro tip: One project, one job. A second job, such as “draft the follow-up email,” is a second project with its own refuse-list. Putting both jobs in one instruction block is how the brief starts writing emails.
Common Mistakes to Avoid
Hiding the rules in the project title. Claude does not read the name or the description. Unsaved instructions are the same as no instructions.
Connecting an inbox so the brief can “send itself.” This job ends on a page you read. Sending is a different job, with a different approval, and it is not this project.
Uploading every note you have ever taken. The next brief will borrow facts from the wrong meeting. Paste the notes for this meeting in the chat. Keep knowledge for definitions that never change.
Going Further
When the brief is dull and correct three times in a row, you can add one line to the output: “Questions I should ask in the room,” limited to gaps already listed under “Not in the notes.” Add it only after the trap in Step 4 still fails closed.
The same shape fits other single jobs: a weekly recap, a support reply that waits for you, a list of unpaid invoices from a file you upload. Each one gets its own project, its own output headings, and a line that says what it must not do. The directory at agentusecases.ai is a pile of jobs other people ran. Pick one job from your own week. Do not install eight bots because a founder did.
Key Takeaways
Claude’s steps: Projects, + New Project, the + button for files, and Set project instructions.
What the current projects do, the free-plan limit, and the newer beta that rolls out through Claude Code.
A directory of jobs people have given agents. Use it to choose one job, not to copy a whole pack of bots.
Sources: