A file that tells Claude who you are does work. Making it is not enough though, where you put it is what counts

This is the prompt sent out in the video. Copy it, paste it into Claude, and the interview starts from there.

Interview me to build my personal context file. Ask one question at a time, 50 total, with checkbox selection. Cover who I am, what I do, how I work, what I hate in outputs, and what great work looks like to me. If my answer is vague, push back and ask for a specific example. When done, compile everything into a single .md file.

What is extra below: where the file goes once the interview is over (the video never says), why the checkboxes may not show up, a corrected version of the prompt that fixes its two snags, and how to tell the file is actually being read.

The video's promise: create the file once and every new chat starts out already knowing you. The second half of that sentence does not happen by itself. When the interview ends you have an .md file, and that is all. When you open a new chat the next day, nobody hands that file to Claude, and Claude starts with no idea it exists.

For the file to do anything, it has to go somewhere Claude reads on its own at the start of every chat. There are three such places and each one reaches a different set of chats.

The two on claude.ai both work on a free account. The only catch with projects is the count: a free account can create at most five.

Using the two together is the cleanest setup. The instructions field in settings goes into every chat you open, the one where you ask for a recipe as much as the one where you draft a work email. So only short rules that hold everywhere belong there. Details like your job, your projects and your examples of good work do more when they sit inside the project where you actually do that work.

Instead of uploading the file you can also paste its contents into the project instructions; the Set project instructions button is on the same page. Both are used in every chat in that project. Putting it in a project has a side benefit too: Anthropic says project content is cached and does not count against your usage limit when reused. Paste the same file into every chat by hand and it counts as part of your message every time.

Claude Code already works this way and reads files named CLAUDE.md on its own at the start of every session. Rules about you that do not depend on a project go in .claude/CLAUDE.md inside your home folder, and apply to every project on your machine.

# Type this in a Claude Code session:
/memory

# Pick the user CLAUDE.md from the list.
# If it does not exist yet it gets created, then opens in your editor.
# Paste in what the interview produced and save.

# Put the file here:
#   macOS and Linux:  ~/.claude/profile.md
#   Windows:          C:\Users\<your-username>\.claude\profile.md

# Then add this line to ~/.claude/CLAUDE.md:
@~/.claude/profile.md

To see that it loaded, type /context in a session and look at the Memory files list. If the file is not there, Claude cannot see it. Anthropic also gives a size guideline: under 200 lines per file. A longer file takes up more room and lowers how well the instructions are followed.

The first is the checkboxes. The feature does exist on claude.ai: instead of waiting for you to type, Claude sometimes offers clickable options, as multiple choice, multi-select or a ranking. But the word in Anthropic's own docs is "sometimes", and they add that it can vary with the type of question. So writing it into the prompt does not guarantee options. If they do not appear the interview is fine, you just type your answer.

The second matters more and sits inside the prompt itself. It asks for checkboxes and also says "if my answer is vague, ask for a specific example". You cannot type an example into a checkbox. Yet the examples are exactly the most valuable part of this file. Asked what bothers you in outputs, a checkbox gets you "too long", a typed answer gets you "opening every reply with 'great question'". Claude can do a lot with the second and almost nothing with the first.

Then there is the number. Each question is its own message, so 50 questions means a conversation of at least 100 messages. Anthropic lists conversation length among the things that affect your usage limit. Making 50 a ceiling instead of a target, and stopping once there is enough, saves both your limit and your patience.

Same idea with the snags fixed. It asks for options only on quick questions, asks for typed answers where an example is needed, makes 50 a ceiling rather than a target, lets you finish early, and ends with two outputs: the full file and a short summary to paste into settings.

We are going to build a personal context file for me. Its purpose: a Claude that reads this file should give me output that fits, without me re-explaining myself in every chat.

Interview rules:
- Ask one question at a time and wait for my answer.
- Ask at most 50 questions. Stop earlier once you have enough. If I type "done", wrap up immediately.
- For quick questions, offer options and always include an "other" option.
- For questions like what bothers me in outputs or examples of good work, do not offer options. Ask me to type the answer.
- If my answer is vague (like "keep it short" or "make it professional"), do not accept it. Ask for a specific example.
- Do not ask for sensitive things like passwords, ID numbers or health details.

Cover: who I am, what I do, how I work, which tools I use, what bothers me in outputs, and what good work looks like to me.

When done, give me two outputs:
1. A single .md file, no longer than 200 lines. Do not transcribe my answers, turn them into rules. Instead of "likes short answers", write checkable sentences like "Open with the conclusion, no intro sentence". Use headings and bullet points.
2. A short summary of the most important rules, 10 bullet points at most. I will paste this into the instructions field in Claude's settings.

Making the sentences in the file checkable is what makes the biggest difference. The example in Anthropic's Claude Code docs puts it well: "use 2-space indentation" instead of "format code properly". The same logic applies to your file. "Make it professional" tells Claude nothing, "no exclamation marks, keep sentences under 20 words" does.

Once it is set up, open a new chat and ask Claude: "What do you know about me, and what should you avoid when writing for me?" If the answer lines up with the rules in your file, it is being read. If you put the file in a project, ask inside that project; a chat opened outside it does not see the project's contents.

One detail: on an account with memory on, part of that answer may come from memory. To check that the file works on its own, add a rule that exists only in the file, for example "end every reply with a one line summary", and see whether it is followed. In Claude Code there is no guessing: type /context and the loaded files are listed.