The feature that stops you re-briefing Claude from scratch every single session. Plain language, and everything checked against Anthropic's own help centre this month.
This is for you if you keep re-explaining the same thing to Claude and you're tired of it. Not you yet? The full reference guide is one click away.
The newsletter you write every week. The client you re-brief every session. The job search. One thing.
Go to claude.ai/projects, hit + New Project, and name it something a tired version of you would recognise at 11pm.
Drop in the doc you keep re-explaining. Then type one line of rules: “I write for women 40+. Plain language. No hype.”
That prompt has Claude ask you the questions instead of you staring at a blank instructions box. Paste it into any chat.
A Project is a workspace inside claude.ai with its own chats, its own uploaded files, and its own standing instructions. Every chat you start inside it begins with that context already loaded. That's the whole thing.
Why it matters: most people pay for Claude, run a 200-message chat, start over from nothing the next morning, and conclude Claude is forgetful. A Project is the difference between briefing a freelancer once and briefing them every time you speak.
Standing identity — voice, role, how you like to be spoken to. Set once in Settings, applies everywhere, doesn't change per Project.
One production: a workflow, a client, a job search. Has its own Context files and its own instructions. Lives for months.
Where you draft badly, try things, get it wrong. Chats inside a Project can't see each other — only Instructions and Context carry across.
The shape travels: ChatGPT calls it Projects or custom GPTs, Gemini calls them Gems, Perplexity calls them Spaces. This guide uses claude.ai because that's where the details are current.
Not every conversation needs a Project. Quick question, one-off task, idle curiosity — open a normal chat, ask, close it. Projects earn their setup time the moment you notice you're doing the same setup over and over.
You write the same kind of thing every week and re-explain the audience every time.
You serve clients and you're tired of pasting the brand brief at the start of every session.
You're job-hunting and want one home for the resume, the target roles, the achievement library.
You're managing something ongoing — a class, a renovation, a diagnosis — with a growing pile of context.
You think “I wish Claude remembered our last conversation” more than once a week.
You open Claude about once a month for one-off questions.
The work you bring is genuinely brand new each time.
You'd rather not commit to a structure. No judgement — it's twenty minutes upfront, and it's twenty minutes you can spend later instead.
There are five doors now, and the names don't help. Answer two or three questions and get pointed at one.
{{ dtHint }}
{{ dtVerdictBody }}
No wrong answers here — you can switch tools later without losing anything you've built.
One-off. Open it, ask, close it.
Same work repeatedly, same context. Set up once, use for months.
A packaged repeatable action — “format this transcript”, “score this resume”. Claude calls it when it fits, or you invoke it directly.
Reads and writes the real files on your computer. Paid plans only. That's guide 02.
Claude Code's ideas in a non-terminal window: agents doing multi-step file work while you watch. Newer, paid, and worth a look once Projects feel easy.
Good news first: you do not need to upgrade to try this. Projects are on every plan, including free. Free accounts get five Projects; Pro, Max, Team and Enterprise get unlimited. Verified against Anthropic's help centre in September 2026 — they move things, so check the link in the references before you quote me.
Individual files cap at 30MB, and there's no limit on how many you add — but a Project has its own capacity meter, and it will refuse files once it's full (somewhere north of 800,000 words). Project capacity and the chat context window are two different budgets.
A regular chat starts empty. A Project starts every chat with your Context and instructions already loaded — more tokens up front, every single time. All-day client work in one heavy Project is how Pro users hit the cap before lunch.
Find the one that looks like your week and steal the structure. The shape is always identical: stock the Project with context you already have, then work inside it for months.
Past issues, voice guide, audience definition — already loaded. Drafts arrive in your voice instead of beige LinkedIn-speak.
Brand brief, scope of work, last three sets of meeting notes. Stop pasting the same context eleven times a week.
Policies, culture deck, the employment law that applies where you operate. Claude handles the offer letter and the awkward email.
The same context tailors the resume, drafts the cover letter, preps the interview, writes the thank-you. The achievement library survives between sessions.
Memo template, board-minutes structure, your exec's voice samples, the quarterly numbers. Fix the sentences, not the bones.
Medical history, insurance details, the IEP, the contractor's estimate. Claude drafts the appeal letter and the question list without the daily re-brief.
About twenty minutes the first time, five every time after. You need a claude.ai account — free is fine — and the doc you keep re-explaining.
Or find Projects in the left sidebar, above your recent chats. If you can't see it, the sidebar is collapsed — toggle it top-left.
Top right. Use a name you'll recognise when you're tired — “Mostly Human” beats “Newsletter v3 FINAL”. The description is for you, not for Claude; skipping it breaks nothing. On Team and Enterprise you also pick visibility here: private, or shared with the org.
This is the file panel — Anthropic calls it Context now, older guides call it “knowledge”. Same thing. Drag and drop, or paste text straight in. Section 7 covers what belongs in there.
The rules of the room: voice, audience, format, hard stops. One paragraph is plenty on day one. Section 8 has a pattern to copy and a prompt that writes it for you.
Not a normal chat — one started from inside the Project view. It already knows everything. “Draft this week's newsletter” is now a complete instruction. That's a working Project.
The fastest way to see what a Project does is to run one prompt twice. Prompt both times: “Write the opening paragraph of this week's issue about AI note-taking tools.”
“In today's fast-paced digital landscape, AI note-taking tools have emerged as a transformational way to unlock productivity. Whether you're a seasoned professional or just getting started, leveraging these solutions can help you optimise your workflow…”
Generic. Every banned word in one paragraph. You'd re-brief from scratch and rewrite the whole thing.
“I tried four AI note-takers this month so you don't have to. Three of them wanted access to my whole calendar before they'd do anything useful. One of them was quietly excellent. Here's the short version, and the part nobody mentions about who owns the transcript…”
Your voice, first try. Lands the audience, skips the banned words, opens on the useful thing. You edit sentences, not structure.
Three to ten well-chosen files beat one giant landfill. Retrieval finds things faster, you burn less of your usage window, and your actual question doesn't drown in irrelevant context.
What makes it sound like you and not a corporate intern. Voice guide, banned words, audience definition, the phrases only you use.
Your best past work. Claude learns your patterns far faster from three real examples than from any abstract rule you write.
The facts and constraints. Brand brief, scope of work, resume, medical history — whatever Claude must not invent.
The shapes you reuse. Memo template, issue structure, cover-letter scaffold. In once, never rebuilt in a chat again.
.md markdown.txt plain text.docx Word.pdf text-based, not scans.csv data30MB per file. No cap on how many — but see “three to ten”, above.
The copy-paste test: open the PDF, try to copy one paragraph. Readable text means Claude can read it. Gibberish means it's a picture — run it through OCR first.
Treat Context the way you'd treat sending the same docs to a careful but brand-new freelancer.
Context is what Claude can look things up in. Instructions are how Claude behaves in the room: voice, format, always, never. One screen of text, hard rules at the top. Not a manifesto — a four-screen instruction set is how the em-dash rule gets buried and ignored.
Paste it into a normal chat, answer honestly for twenty minutes, and you'll have instructions you would never have written from a blank page. Approach adapted from Ruben Hassid's voice-interview prompt (see references).
Five real failures, all fixable, all coming from the same place: the Project knows less than you think it does, or it knows the wrong thing.
Projects pile up: the newsletter you don't write anymore, the finished client engagement, the search that ended in a job. On free they also keep eating your five slots. Once a quarter: open the list, be honest about what you've touched in ninety days, export the Context files you want to keep — chats inside a Project go away with it — then delete or archive the dead ones. One caveat on archiving: on the free plan an archived Project can become unavailable rather than merely tidy, so export first.
APA 7th edition. This redesign is derived from Deb Haas's original guides, with facts re-verified against Anthropic's documentation in September 2026.
Anthropic. (2026). How can I create and manage projects? Claude Help Center. https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects
Anthropic. (2026). Retrieval-augmented generation (RAG) for projects. Claude Help Center. https://support.claude.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects
Anthropic. (2026). What are projects? Claude Help Center. https://support.claude.com/en/articles/9517075-what-are-projects
Haas, D. (2026). Claude Projects for humans (Version 1.0) [Web guide]. Mostly Human. https://claudeprojectsforhumans.netlify.app/
Hassid, R. (2026, January). The “taste interviewer” prompt [Post]. LinkedIn. https://www.linkedin.com/posts/ruben-hassid_how-to-make-ai-sound-exactly-like-you-forever
Meehan, C., & Stelzner, M. (2025, December). Using Claude Projects to develop quality content. Social Media Examiner. https://www.socialmediaexaminer.com/using-claude-projects-to-develop-quality-content/
Everything you'd otherwise learn the hard way — in an order that makes sense for someone who has never written a line of code. Read it top to bottom the first time; each part sets up the next.
Everything else is detail. These two turn “an AI is loose in my folders” into “an assistant that shows me its work”.
Claude tells you exactly what it intends to do and touches nothing until you say go. Reach for it before any task involving more than one file.
/rewind is a real undoIt rolls back the conversation and the file changes to an earlier checkpoint. Use it freely. It is the reason you can afford to experiment.
Ten seconds of duplicating the folder you care about has saved more people more pain than any setting in this guide.
It is not a chatbot, and that's the part nobody leads with. In claude.ai, Claude talks to you: words in, words out. Claude Code reads the files on your computer, writes and edits them, and runs commands — inside your actual folders, not a demo sandbox.
Regular Claude is a brilliant advisor on the other side of a window. Claude Code is the same advisor in the room with you, able to pick things up and move them.
Whole folders, transcripts, exports, spreadsheets. It can answer questions about material far too long to paste into a chat.
Creates files, edits documents in place, restructures folders. This is the powerful part and the part that needs plan mode.
Builds a site, converts a batch of files, runs a script. You don't need to know the command — you need to know what you want.
You don't need to know how to code. You need to know what you want done. “Rename everything in this folder with today's date.” “Read this transcript and pull out every action item.” “Build me a one-page site for this offer.” That's the level of instruction that works — and being specific matters more than being technical.
Installing is free; using it requires Pro, Max, Team, Enterprise, or a pay-per-token API key from the developer console. The free claude.ai plan does not include it. Nobody tells you this upfront, and it's annoying, so: now you know.
Fine for personal projects and learning. You'll notice the rolling five-hour window if you work in one long session all afternoon.
For multiple projects, long sessions, most of the day. Flat monthly rate — unlike per-call API billing, you know what you're paying.
Long sessions eat context fast. Working in one never-cleared conversation is the single biggest reason people hit limits early. /context and /clear in section 6 are the fix, and they're free.
Every tutorial is written for developers. Here's what it looks like in the hands of someone who isn't one.
Hand it the data exports from LinkedIn, Facebook and ChatGPT and have it sort every piece of your writing into one tagged, searchable vault. Days of work, one afternoon.
It reads a message export, matches names to the right contact files, and appends notes to dozens of records without you opening one of them.
Sales pages, resource pages, landing pages — complete builds from a description. You never touch HTML unless you want to.
From a rough idea to a structured deck: it thinks through flow, proposes a shape, writes the slide content. Good enough for client and workshop material.
Messy folders, inconsistent names, things saved in the wrong place. It can audit and restructure at a scale that would make you cry by hand.
Tell it what you're after — action items, one person's comments, the decisions — and it pulls exactly that out of a pile of meeting notes.
It is not magic. It makes mistakes, occasionally does something you didn't ask for, and can loop on a fix that creates the next problem. Treat it as delegating to a very capable person who needs clear direction and whose work you check. Never point it at anything with live credentials or production data while you're still learning what it does.
There are more ways to install this than anyone needs. Two are worth your time: the desktop app if the terminal makes you tense, or the native one-line installer if you'd rather have the full thing. Both work on Mac and Windows.
Anthropic rebuilt the desktop app in 2026 around exactly this: a proper window for agentic file work, with a sidebar of sessions, parallel work isolated from each other, and visual diffs you approve before anything is applied. Mac and Windows only — Linux users take the terminal path.
Download it from claude.com/download and install it like any other app.
Sign in with your Pro or Max account and open the Code tab.
Point it at a folder and say what you want. Review the diff, then apply.
Open Terminal (Cmd + Space, type Terminal) and paste this. No Node, no other software first. It auto-updates itself afterwards.
curl -fsSL https://claude.ai/install.sh | bash
Then type claude, log in through the browser, and you're at the > prompt. Homebrew instead: brew install --cask claude-code — but Homebrew doesn't auto-update, so run brew upgrade claude-code now and then.
Claude Code runs natively on Windows now — no WSL required. Open PowerShell (not Command Prompt) and run:
irm https://claude.ai/install.ps1 | iex
Close and reopen the terminal, then type claude. WinGet instead: winget install Anthropic.ClaudeCode. Installing Git for Windows is optional but recommended — with it, Claude Code uses Git Bash for shell commands.
The npm package still works (npm install -g @anthropic-ai/claude-code, needs a current Node) but it isn't the path Anthropic tests first and it won't update itself. Most old tutorials you'll find online assume it — that mismatch is where a lot of “nothing works” comes from.
If you want your files visible alongside Claude, install VS Code and add the official Claude Code extension from the Extensions panel. Good middle ground; not necessary on day one.
Close and reopen the terminal first — that solves most “command not found”. Then run claude doctor, which reports how it was installed, the version, and what's wrong. Permission errors on Windows usually mean PowerShell needs to be run as Administrator.
If you implement one thing from this guide, make it this. CLAUDE.md is a plain text file that tells Claude Code how to work with you. It's read at the start of every session, so you stop re-explaining yourself every time you open a new one. It is the Claude Code equivalent of a Project's instructions.
About you: how you think, how you want to be spoken to, what Claude must never do without asking. Read every session, in every folder.
Mac: ~/.claude/CLAUDE.md · Windows: the same path inside your user folder.
Lives in a specific folder and describes that work: the goal, the files involved, the rules. Inside that folder both files are read, and this one wins any conflict.
Fastest way to make one: run /init in the folder and edit what it generates.
CLAUDE.md is guidance, not a guard rail. It shapes behaviour reliably; it does not enforce anything. For rules that genuinely must hold — don't touch this folder, don't publish anything — say it in the session too, and keep plan mode on.
Easiest way to create it: start a session, paste the filled-in template, and say “save this as CLAUDE.md in my home directory”. Claude Code puts the file in the right place for you.
Everyone mentions these; nobody explains them. Grouped by what you actually need them for. The set changes with releases — /help inside a session is always the current truth.
/clear matters more than it sounds. Long sessions accumulate context, and Claude starts carrying earlier parts of the conversation into what you're asking now, muddying both. When in doubt, clear it and restate the goal. It isn't failure. It's hygiene.
The point is making it slow down and confirm while you're still learning what it's capable of. Once you trust the rhythm, you'll drop the training wheels on small tasks and keep them on for anything that touches files you'd hate to lose.
APA 7th edition. Commands and install paths re-verified against Anthropic's documentation in September 2026; the slash-command set moves with releases, so trust /help over any page including this one.
Anthropic. (2026). Claude Code: Set up Claude Code. Claude Code documentation. https://code.claude.com/docs/en/setup
Anthropic. (2026). Claude Code: Slash commands. Claude Code documentation. https://code.claude.com/docs/en/slash-commands
Anthropic. (2026). Claude Code: Troubleshooting. Claude Code documentation. https://code.claude.com/docs/en/troubleshooting
Haas, D. (2026). The guide Anthropic should have written (Version 1.0) [Web guide]. Mostly Human. https://claudecodeforhumans.netlify.app/
Kelly, C. (2026). Claude Code troubleshooting: Five fixes that solve ninety percent of problems. Claudefast. https://claudefa.st/blog/guide/troubleshooting
r/ClaudeAI. (2026). Claude Code for non-coding projects: A complete guide [Online forum post]. Reddit. https://www.reddit.com/r/ClaudeAI/comments/1q403xw/