AI Guides › Step-by-step guides
Five Claude Projects for Recurring Admin, Built in a Week
By Nigel Guy · 6 min read
Most people open a fresh chat, type a paragraph about who they are and what they need, get a decent answer, and repeat the whole paragraph tomorrow. It feels productive because each answer is good. The cost is hidden: you are the memory, and you pay for it every single time.
A Claude Project moves that paragraph into a place Claude reads at the start of every chat. The usual mistake is to build one giant "work" project and fill it with everything. It then knows a little about all of it and a lot about none.
The rule: one Project per recurring job, with written instructions, two or three reference files and one real task handed over on the day you build it.
Before you start
- Account. Projects are available on every plan including Free, but Free is capped at five projects (per Anthropic's help centre). That is exactly enough for this guide.
- Cost. Free is £0. Pro is $20 a month ($17 a month billed annually), about £15 a month at time of writing; Max starts at $100 a month, about £75 at time of writing. You do not need Max for this. Check the £ price at checkout, as exchange rates and tax vary.
- Paid-plan bonus. On paid plans, Claude switches on retrieval ("RAG") mode automatically when a Project's knowledge nears the context limit, which lets it hold considerably more. On Free you need to keep files short.
- Your raw material. Three to five real examples per job: a past email you were pleased with, last week's notes, a template. Anonymise anything that identifies other people.
- Time. About 20 minutes per Project.
The five Projects
| Project |
Recurring job |
Starter files |
| 1. Reply Desk |
Drafting routine emails and messages in your voice |
Three of your own sent emails, your sign-off and tone notes |
| 2. Meeting to Actions |
Turning notes or transcripts into decisions, owners and dates |
One past set of good minutes |
| 3. Weekly Review |
A Friday summary of what moved and what is stuck |
Your current priorities list |
| 4. Chase Desk |
Polite payment, reply and deadline chasers |
Your payment terms, your usual wording |
| 5. Document Checker |
Checking a draft against your own standards before it goes out |
Your style notes or checklist |
Build the first three in the order shown: they use material you already have. Projects 4 and 5 tend to repay the effort quickest, because each replaces something you currently put off or fix late.
Step 1 — Create the Project
- Go to claude.ai/projects, or choose Projects from the sidebar or account menu.
- Click New project (top right).
- Give it a plain name and a one-line description. Per Anthropic, Claude does not read the name or description, so they are for you.
Step 2 — Let Claude interview you for the instructions
Do not write instructions from a blank page. Open a normal chat and paste this.
Role: you are a workflow designer helping me set up a Claude Project.
Context: the Project is called [PROJECT_NAME]. Its job is to handle [RECURRING_JOB]. I do this job about [HOW_OFTEN]. My audience or counterpart is [WHO_IT_GOES_TO]. My tone is [TONE_IN_THREE_WORDS].
Goal: a Project I can open cold and use immediately, with no need to explain my situation again.
Steps:
1. Ask me up to eight short questions about what you would need to know to do this job well: preferences, recurring facts, things to avoid, and what a finished result looks like. Ask them one batch at a time and wait for my answers.
2. Once I have answered, write the project instructions in second person ("You are...").
3. List the exact files or text I should add on day one.
4. Propose one real task for me to give the Project this week.
Output format: three labelled sections: Instructions, Day-one files, First task. Keep the instructions under 400 words.
Constraints: do not invent facts about me or my business. If an answer is missing or vague, ask again rather than guess. Do not include anything confidential in the instructions unless I say so.
Self-check before you finish: confirm the instructions state the job, the tone, the output format, at least two things to avoid and what to do when information is missing.
Fill in the five square-bracket items; answer Claude's questions honestly and briefly.
Step 3 — Paste the instructions into the Project
- Inside the Project, click Set project instructions.
- Paste the Instructions section from Step 2 and save. Anthropic states Claude uses these for all chats in the Project.
- Add one line at the end: "If you are missing information you need, ask me before drafting."
Step 4 — Add the day-one files
- On the Project page, click the + in the knowledge panel.
- Upload or paste the files Claude listed: your sample emails, template, priorities list.
- Keep each file focused. A tidy one-page "how I like things" note beats a folder dump.
Step 5 — Hand it the real task
Start a new chat inside the Project and use the first task Claude proposed. Use live work, not a test. You learn more from a real email you must send than from a made-up one.
Step 6 — Fix the instructions, not the chat
When the output is wrong, resist correcting it in the chat and moving on. Ask: "What should the project instructions say so you get this right next time?" Paste the answer into the instructions. That edit is what makes Project number two faster than number one.
Check it worked
- Open a new chat in the Project and give it a one-line request. It should match your tone and format without being told.
- Open a normal chat outside the Project with the same request. The difference is what you built.
- By Sunday, each Project should have been used once on real work.
What to skip
- One mega-Project. Separate jobs get separate instructions.
- Uploading everything. More files dilute attention and hit limits sooner.
- Customer or staff personal data. Leave it out unless you have checked your obligations and your plan's data terms.
- Sharing. Anthropic lists project sharing as a Team and Enterprise feature; on Free, Pro and Max it is not available.
- Fancy tricks. Do not bother with elaborate personas. A role, a job, a format and rules are enough.
Guardrails
- Context does not carry between chats by itself. Anthropic says information in one chat is not shared with others in a Project unless it is in the knowledge base or memory. Memory is a separate setting with its own per-Project space; check yours under settings, and use an incognito chat for anything sensitive, since incognito bypasses memory.
- A Project will not act for you. It drafts and structures. It does not send emails, chase invoices or update your calendar. You press send. I could not verify any scheduled or automatic running of Projects in the help centre, so do not rely on one.
- Read every draft before it leaves. Claude can state wrong dates, amounts or names with confidence, especially when your notes are thin.
- Review the instructions monthly. Anything stale in them will be applied faithfully.
- Plans and prices change. Check the live pricing page before paying.
Sources
All 751 AI guides · JulieMango plans from £17/mo