AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
Most people treat a Claude Project as a folder: they drag in every document they own, write nothing in the instructions box, and wonder why the answers sound generic. A project only earns its keep when the instructions say who you are and what good looks like, and the files are a short, named, current set. Dumping files in feels like progress while quietly making retrieval worse.
The rule: one project per area of your life, one page of pinned instructions, and a short upload list you can name from memory.
Build these five, or swap in your own. The test is whether you keep explaining the same background in fresh chats.
| Project | What it holds | Pinned instruction in one line |
|---|---|---|
| Content | Your voice samples, audience notes, a "never say" list | Write in my voice; draft first, ask second |
| Business | Offer summary, pricing, customer notes, brand rules | Act as a sceptical operations partner |
| Finances | Your own budget summary and tax-year notes | Explain, never advise; flag what a professional should check |
| Research | Reading notes, your questions, source standards | Cite which file each claim came from |
| Personal admin | Household facts, routines, key dates | Keep answers short and list next actions |
For Finances, upload summaries you have written rather than raw statements. Anything sensitive that you would not email to a stranger stays out.
Inside the project, open a new chat and paste this. It interviews you, then drafts the instructions and an upload list.
You are helping me set up a Claude Project so it works as a long-term assistant for one area of my life.
Project area: [AREA, e.g. content, business, finances, research, personal admin]
What I usually want from it: [TYPICAL TASKS]
How I like answers: [LENGTH, TONE, FORMAT PREFERENCES]
Things I never want: [BANNED HABITS OR TOPICS]
Step 1: Before writing anything, ask me up to five short questions about what you would need to know to be useful without me re-explaining each time. Wait for my answers. If any input above is blank or vague, ask about it.
Step 2: Write a set of project instructions of no more than 350 words, covering: your role, my context, what a good answer looks like, the default output format, and your guardrails.
Step 3: List exactly which files I should upload, in priority order, and name the kinds of files that would only be noise.
Step 4: Mark every point where you guessed rather than used something I told you.
Self-check before answering: confirm each instruction is specific and testable, nothing contradicts anything else, and no personal data beyond what I supplied has been invented.
Fill in the four bracketed lines. Answer the questions, then read the draft critically; you are the editor.
Content shares the context window with your instructions and your conversation. When knowledge approaches that limit, paid plans switch on RAG mode automatically, which the help centre says expands capacity by up to 10x. You will see an indicator, and Claude will use a "project knowledge search" tool instead of reading everything at once. Two consequences: retrieval can miss a passage a full read would have caught, and a cluttered knowledge base gives it more to miss.
Start a fresh chat in the project and ask three things:
If any fails, edit the instructions or remove a muddy file, then retest.
Nothing refreshes itself. Prices change, your offer shifts, a file is superseded, and Claude keeps trusting the old version. Put a recurring entry in your calendar, monthly for fast-moving projects and quarterly for the rest: replace changed files, delete dead ones, and reread the instructions. Memory also works per project: the help centre says each project has its own separate memory space and summary. Memory settings live under Settings > Memory, and whether memory is available on the free plan is described inconsistently across Anthropic's pages, so check your own settings rather than assuming.
It will not make Claude correct. It reads your files and may still summarise them wrongly. It does not watch your inbox or update itself. It is not a replacement for a notes app you control, and it is not independent financial, legal or tax advice.