AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
Most people open Claude, type a one-line question, get a competent but generic answer and conclude the tool is average. It feels fine because the answer is not wrong. It is just written for nobody in particular, because Claude knows nothing about you, your work or what "good" looks like for you.
The rule: tell Claude everything once, in a project, before you ask it for anything that matters.
That means three moves, in this order: a messy brain dump, a project that holds the instructions Claude drafts from it, and a small set of files that back those instructions up. Nothing here needs technical skill.
Open a new chat (not a project yet). Paste the prompt below, then type everything: what you are working on, what is stuck, what you keep avoiding, what worries you, what you want by a particular date. Messy and out of order is correct. Do not edit as you go.
You are a thoughtful chief of staff meeting me for the first time. I am about to give you an unfiltered dump of what is on my mind. It will be disorganised and repetitive, and that is deliberate.
Rules for this turn:
- Do not summarise, tidy or advise yet. Just read it all.
- Wait until I write "DONE" before you respond.
When I write "DONE", reply with:
1. Three to five themes, each in one sentence, grouping items that are really the same problem described differently.
2. What I seem to be avoiding or not saying, based only on what I wrote. Mark anything you are inferring as an inference.
3. Up to five questions you would need answered to help me properly. Do not guess answers.
Keep the whole reply under 400 words, in plain British English, no flattery.
Before sending, check that every point traces back to something I actually wrote.
Here is my dump:
[EVERYTHING: WHAT I'M WORKING ON, WHAT'S STUCK, WHAT I KEEP PUTTING OFF, WHAT WORRIES ME, WHAT I WANT BY [DATE]]
Fill in the square brackets with your own mess. Then type DONE, read the reply and answer its questions in the same chat. Correct anything it got wrong; this is the cheapest moment to do it.
In the same chat, send this:
You are helping me set up a Claude project for [AREA OF MY LIFE OR BUSINESS].
Using everything I have told you in this conversation, draft the project instructions that every future chat in this project will load. Aim for something I can paste straight in.
Structure:
1. Who I am (role, situation, relevant constraints), in a short paragraph.
2. What I am working on and what a good outcome looks like, with dates where I gave them.
3. How I want you to talk to me (tone, length, spelling, what to avoid).
4. Standing rules: things to take into account every time without me repeating them.
5. When to ask me a question rather than assume.
Constraints: use only facts I have given you. Where something is missing, put a line in [SQUARE BRACKETS] for me to fill in. Do not invent details. Keep it under 500 words.
After the draft, list exactly which files I should upload to make you more useful here, and why each one matters. Then check your draft for anything I did not actually say, and remove it.
Now create the project. At time of writing, Anthropic's help centre gives these steps:
Read the draft before you paste it. If it says something untrue about you, fix it now, because it will colour every future chat.
On the project page, add files to the project knowledge area (the help centre describes it as the right-hand side of the project). Claude uses these as context in chats inside that project.
Good first files, depending on your area:
| If your project is about | Upload |
|---|---|
| A business | Your own one-page description of it, a price list, two examples of your best writing, your customer list headings (not personal data) |
| A job search | Current CV, two job adverts you like, notes on what you want next |
| Study | Syllabus, assessment criteria, your own notes |
| Personal admin | A summary of deadlines, a budget outline with no account numbers |
Keep it to a handful of documents at first. Short, current files beat a dump of every document you own. Paid plans switch on retrieval automatically when a project nears its context limit, according to the help centre, but the free route to better answers is still curation.
Start every chat about this area inside the project. Open chats for separate tasks rather than one endless thread. Tell Claude when something changes: "Update: the deadline moved to [NEW DATE]." Then amend the instructions when the same correction comes up twice.
Open a fresh chat inside the project and ask something you would normally have to explain: "What should I do first this week, and why?" A working setup gives you an answer that names your real situation, uses your preferred tone and does not ask you who you are. If it still sounds generic, either the instructions are thin or the files do not say what you think they say. Compare against the draft from Step 2.
Project instructions are a living document, not a one-off.
The trap is writing the instructions in the abstract ("be concise, be helpful, act as an expert") instead of from your own mess. Generic instructions produce generic answers. The brain dump matters because it is where your specific constraints, worries and goals come from. The second trap is uploading everything: more files are not better, and stale ones can pull answers the wrong way.