Claude Projects: The Complete Setup Guide

A normal Claude chat starts from zero every time: you paste the background, explain the tone, upload the same files again. A Project ends that. A Project is a standing workspace inside Claude with its own instructions (a brief that applies to every chat in it) and knowledge (files Claude can always see). Set one up once, and every conversation inside it starts already knowing your context.

Projects are available on every Claude plan — free accounts get up to 5 Projects; paid plans (Pro, Max, Team, Enterprise) get unlimited, with sharing and collaboration on paid plans (per Anthropic’s plan comparison, checked 18 October 2026). This guide walks through a proper setup in about ten minutes, the instruction-writing rules that make Projects actually work, and five blueprints you can copy.

Contents
– What a Project is (and what it isn’t)
– Step 1: Create the Project and name it like a job
– Step 2: Write instructions that hold
– Step 3: Build a knowledge base that stays sharp
– Step 4: Work inside the Project properly
– Five starter blueprints
– Free vs paid: what actually changes
– Common mistakes
– Frequently asked questions
– The bottom line

What a Project is (and what it isn’t)

A Project has three parts:

• Instructions — a standing brief. Every chat in the Project follows it: your role, your audience, your formats, your rules.
• Knowledge — files you upload once (PDFs, documents, spreadsheets, text files, up to 30MB each). Claude can reference them in every chat in the Project.
• Chats — ordinary conversations, except they all inherit the instructions and knowledge above.

A Project is not a folder for finished work, and it is not memory. Claude’s memory carries your preferences across all chats; Project instructions are scoped — they apply only inside that Project, which is exactly what you want when “write like a lawyer” applies to one client and “write like a friend” to another. Keep universal habits in your profile preferences; keep job-specific briefs in Projects.

Step 1: Create the Project and name it like a job

1. Open claude.ai (or the desktop/mobile app) and click Projects in the left sidebar.
2. Click New project.
3. Name it after one job with one audience: “Client: Acme — content retainer,” “Q3 board reporting,” “Job hunt 2026.” Not “Work,” not “Claude stuff.”

The naming rule matters more than it looks: a Project that tries to write proposals, answer support tickets and draft social posts does all three badly, because its instructions blur. One Project per recurring job is the whole system.

Free accounts can hold up to 5 Projects, which is enough for your five most repetitive jobs. Paid plans remove the cap and add project sharing — a team can work inside the same brief and knowledge base.

Step 2: Write instructions that hold

Click Set project instructions and write the brief. Good instructions are short, specific and written as orders, not wishes. Include:

• Who you are and what this Project is for — one or two sentences.
• The audience — one sentence, in plain words.
• Output formats — “Draft emails under 150 words.” “Give me three options for headlines, one recommended.”
• Style rules — three or four non-negotiables, pulled from your real style guide if you have one.
• A forbidden list — words, phrases, topics or claims Claude must not use here.
• One “never” line — the single failure you care most about preventing (“Never invent a statistic; if a number isn’t in the knowledge files, say so”).

Two discipline rules from experienced users: keep instructions under roughly 400 words — Claude weighs instructions broadly equally, so padding dilutes your most important rules — and write them in the imperative (“Use British spelling”) rather than the descriptive (“we usually tend to use British spelling”).

Step 3: Build a knowledge base that stays sharp

From the Project page, add files to the knowledge panel. What belongs there:

• The 5–10 documents you reach for most in this job: the style guide, the master price sheet, the product spec, the template, the brand examples, last quarter’s best report.
• Examples of your best finished work — two or three pieces that sound like you. This is the highest-value upload for writing Projects; Claude matches demonstrated style better than described style.

And the rules that keep it useful:

• Quality beats quantity. Small knowledge bases are read in full. Once a Project’s files grow large, paid plans switch to retrieval — Claude pulls the most relevant passages per answer rather than reading everything — so one bloated Project quietly becomes a less reliable one. Split big topics into separate Projects instead.
• Delete stale files immediately. An outdated price sheet in the knowledge base produces outdated prices in every answer, and Claude cannot flag which of your files you stopped believing.
• Redact before uploading. No passwords, personal data or confidential material that doesn’t need to be there for the work.

Step 4: Work inside the Project properly

• Start a fresh chat for each new task. Every new chat still has the full instructions and knowledge — you lose nothing by starting clean, and you avoid the cost of a long history being re-processed with every message.
• Move existing chats in. From your chat history you can select past conversations and move them into the Project to consolidate scattered work.
• Star your active Project so it stays pinned at the top of the sidebar; archive finished ones to keep the list honest.
• When output drifts, fix the brief, not the chat. If Claude breaks the same rule twice, the instruction is missing or ambiguous. Edit the instructions once and every future chat improves.

Five starter blueprints

1. The content desk. Instructions: your audience, voice rules, three best articles as knowledge, your SEO basics. Every draft starts on-brand; use it for posts, newsletters and scripts.

2. The client workspace. One Project per client: their brief, brand guide, past deliverables and meeting notes as knowledge. Status updates, proposals and emails all draw on the same facts.

3. The support inbox. Your help-centre articles, tone guide and escalation rules as knowledge; instructions to draft replies in your support voice and flag anything outside policy. Paste the ticket, get a sendable draft.

4. The research file. A topic you track over months: key papers, data tables and your own notes as knowledge; instructions to always cite which knowledge file a claim came from. New chats brief you like an analyst who’s read the file — because it has.

5. The job hunt. Your CV, the job descriptions you’re targeting and your cover-letter exemplars as knowledge; instructions for tailoring, interview prep and follow-up emails in one consistent professional voice.

Keeping Projects healthy over time

A Project is a living brief, and it decays the same way a shared drive does — by accumulation. Three maintenance habits, five minutes a month each, keep it sharp:

• A monthly knowledge audit. Open the knowledge panel and delete anything you no longer believe: superseded price sheets, last quarter’s roadmap, the draft that lost. If two files disagree, Claude has no way to know which one you trust — you have to make the disagreement impossible.
• An instructions changelog. When you edit the instructions, note what changed and why at the top of the file or in your own notes. When output quality shifts, you’ll know whether the model changed or you did.
• One Project, one owner. On shared (paid) Projects, agree who curates the knowledge base. Five people uploading freely recreate the mega-Project problem inside a shared brief within a month.

Treat the Project like a good employee handbook: short, current, and edited on purpose — never a drawer everything gets thrown into.

Free vs paid: what actually changes

For a solo user with a handful of recurring jobs, the free allowance of 5 Projects is not a teaser — it covers the system. Paid becomes necessary when you run out of Project slots, need a team inside the same brief, or hit the free usage limits mid-work.

Common mistakes

• One mega-Project for everything. Blurred instructions, noisy retrieval, mediocre everything. Split by job.
• Instructions written as a diary. Background essays bury the orders. Brief first, context second, keep it tight.
• A knowledge base you never prune. Stale files are wrong answers waiting to happen.
• One endless chat. Long threads cost more per message and drift. New task, new chat — the Project remembers what matters.
• Expecting the Project to fix a bad brief. If you can’t state the format and the rule, Claude can’t follow them. The setup work is the work.

Frequently asked questions

Are Claude Projects free?
Yes. Anthropic’s plan comparison lists Projects on the Free plan, capped at 5, with unlimited Projects on paid plans. Sharing and collaboration are paid-plan features.

What’s the difference between a Project and a normal chat?
A normal chat starts blank and forgets when it ends. Every chat inside a Project automatically follows the Project’s instructions and can see its knowledge files — no re-pasting context, ever.

How big can knowledge files be?
Up to 30MB per file, across common document, spreadsheet, presentation and text formats. Small sets are read in full; on paid plans, very large knowledge bases switch to retrieval, where Claude pulls the most relevant passages into each answer.

Do Project instructions override my general preferences?
They’re scoped: profile preferences shape all your chats, Project instructions shape only that Project’s chats. Keep universal rules in your profile and job-specific briefs in Projects, and the two rarely fight.

Can my team share a Project?
Yes, on paid plans — project sharing and collaboration are listed as paid features in Anthropic’s comparison. Free Projects are yours alone.

Projects vs Custom GPTs — which is better?
They solve different problems. A Project is a place you work — a brief plus your files around your own ongoing jobs. A Custom GPT is a tool you build and can share publicly. For personal and team workflows, Projects are faster to set up and maintain; for publishing a reusable bot to an audience, that’s the GPT pattern.

Should I put client-confidential files in a Project?
Only material the work genuinely needs, and only within your organisation’s data policy. Redact what isn’t required, remember that knowledge files shape every answer in the Project, and use the plan tier whose data terms your work demands — Anthropic publishes plan-level controls (including training opt-out on its pricing page), but your own compliance obligations are yours to check.

The bottom line

A Project is the difference between an assistant you brief every morning and one that turns up already briefed. Ten minutes of setup — a job-shaped name, a tight instruction brief, a pruned knowledge base — pays you back on every chat afterwards, and the free plan’s five Projects are enough to run the system properly.

Your next step: pick your single most repetitive task this week, build its Project from Blueprint 1, 2 or 3, and run tomorrow’s instance of that task inside it. If the output doesn’t need less editing than today’s did, tighten the instructions — that feedback loop, not the tool, is the skill.

Plan details checked against Anthropic’s official pricing page comparison table on 18 October 2026; interface mechanics cross-checked against Anthropic’s documentation as reported in October 2026 setup guides. Features and limits can change — Anthropic’s help centre is the source of truth.