How to use Claude Projects for content work in 2027

Key takeaways:

• A Claude project is a workspace with its own instructions, knowledge files, chat history and memory, so your brand guide and standing corrections load into every chat instead of being re-pasted.
• Projects are on every Claude plan, including Free, where Anthropic caps you at five. Sharing a project with colleagues needs a Team or Enterprise plan.
• The setup that pays is one project per recurring content type, an instruction block that fits on one screen, and only the files that type needs.
• A project is the wrong tool for one-off research, work that depends on this week's sources, and jobs that should run on a schedule.

Most content teams use Claude like a search bar: open a fresh chat, paste the brand guide again, paste the brief again, correct the same five mistakes again. A project ends that ritual. It's a workspace where the context lives once, and every conversation inside it starts from what you've already taught it.

This guide covers how to use Claude Projects for content work: what a project holds, what each plan gets, the setup steps, a worked instruction block you can adapt, how to split projects across a team's work, and where a project is the wrong tool.

At busyless I build AI content pipelines for clients, and every one of them rests on the same habit: a rule gets written down once and read on every run, rather than retyped into a chat.

What a Claude project is

Anthropic's support documentation defines projects as self-contained workspaces with their own chat histories and knowledge bases. Inside one, you upload documents, give Claude context, and hold focused chats. So what is a Claude project in practice? It's the container that stops your second chat from forgetting what you told the first.

Projects aren't new. Anthropic launched them on June 25, 2024, for Pro and Team customers, and its launch announcement gave each project a 200,000-token context window, which it described as the equivalent of a 500-page book. The same post quoted North Highland, a consultancy, saying Claude helped its team finish content creation and analysis tasks up to 5x faster. That's a customer's claim, but it shows where projects were aimed from day one.

Each project carries five pieces, and the table shows which of them Claude actually reads.

The "No" in the first row catches people out. Anthropic's help center says Claude has no access to a project's name or description, so a description like "Use our casual voice" does nothing. Anything you want Claude to follow belongs in the instructions or the knowledge files.

The chats row matters too. Anthropic notes that context isn't shared across chats within a project unless you add it to the project knowledge, so don't count on a correction from Tuesday's chat reaching Wednesday's. Project memory may pick some of it up, but only an instruction line or a knowledge file is certain to load.

What each plan gets

Projects are available on every plan, but the limits differ, and the differences matter for a content team. The prices below come from Anthropic's pricing page as of November 2026.

Two limits sit outside the table. Paid plans switch a project to retrieval mode automatically when its knowledge approaches the context limit, which Anthropic says expands capacity by up to 10x and describes as a paid-plan feature. And only Team and Enterprise members can share a project with colleagues, giving each person "Can view" or "Can edit" access.

For a solo writer, Free's five projects are enough to test the idea on your two most repeated jobs. For a team, the sharing row is the deciding one: without it, every writer builds and maintains a private copy of the same project.

How to use Claude Projects: the setup

The steps below are the whole procedure for a content project, using the menus as Anthropic's help center describes them. Budget an afternoon for the first one, most of it spent on step 7.

1. Go to claude.ai/projects, or click "Projects" in the left sidebar, then click "+ New Project".
2. Name the project for the content type ("Client newsletters", not "March stuff"), and use the description to tell colleagues what it's for, since Claude never sees it.
3. On a Team or Enterprise plan, choose whether to keep it private or share it with your organization.
4. Click "Set project instructions", paste your instruction block, and click "Save instructions".
5. Click the "+" button in the project knowledge panel and upload the brand guide, the ICP one-pager and two or three approved examples.
6. If a past chat produced your best draft of this content type, open its dropdown menu and choose "Add to project" so it lives with the rest.
7. Run one real piece of work inside the project and correct the output once, thoroughly.
8. Turn every correction from step 7 into an instruction line, so the next chat starts with the fix already learned.

Teams that skip step 8 are the ones still making the same correction in month three. That's how to use Claude Projects day to day: the project is the container, the instruction block is the training, and the knowledge files are the source of truth, so most of what follows is about getting the instruction block right.

The instruction block that does the work

The instruction block is where a project earns its keep, and the difference between useful Claude project instructions and decorative ones is specificity. "Write in a friendly, professional tone" governs nothing. "Second person, contractions, sentences under 25 words" governs every line.

A useful block covers five things: who the reader is, which sources Claude may use, the voice rules, the structure of the output, and the quality bar that makes a draft shippable. It fits on one screen, because a block nobody rereads stops being maintained, and an unmaintained block drifts away from what the team actually wants.

Here is a worked example for a B2B SaaS blog project. The company is invented, so swap in your own reader, files and rules.

You draft blog articles for Acme's content team.

Reader: heads of finance at SaaS companies with 50 to 500 staff.
They know accounting well. They don't know our product.

Sources: use only the files in project knowledge and anything
pasted into this chat. If a claim needs a number you don't have,
write [SOURCE NEEDED] instead of estimating.

Voice: follow brand-voice.pdf. Second person, contractions,
sentences under 25 words. Banned words are in kill-list.txt.

Structure: H1, three summary bullets, then H2 sections of
150 to 300 words. Open every section with a full sentence.
End with three FAQs.

Before drafting: send an outline and the three files you plan
to use, then wait for my approval.

Shippable means: every number has a source, every section makes
one point, and no paragraph copies an approved example.

Each line exists because of a failure it prevents. The [SOURCE NEEDED] rule stops Claude inventing a statistic to fill a gap, one of the costliest mistakes a published article can carry. The outline-first rule catches a wrong angle in two minutes instead of after a 1,500-word draft. The last line turns "is this good?" into three checks anyone on the team can run.

Notice what the block leaves out: the brand guide's actual content and the full banned-word list. Both live in project knowledge, and the instructions point to them by filename. Duplicated rules drift out of sync the first time someone updates one copy, which is the lesson behind keeping the system on disk: every rule lives in one file, never in a chat.

Keeping a project current

Claude project instructions compound only if they change with use. My rule is simple: when you correct the same mistake twice, the correction becomes an instruction line. That's how a project improves instead of just storing files.

I saw the same loop while building the busyless site with Claude Code, a build written up among my case studies. Every time we hit a library quirk, Claude wrote the fix and added a note to a "things to watch for" file, so the next session read the warning before the mistake could repeat. A project's instruction block works the same way, just inside claude.ai rather than a code folder.

Knowledge needs pruning as well as adding. Run this check once a month on every project your team relies on:

• Delete any approved example that no longer matches how you write today.
• Replace the brand guide and ICP files whenever the source documents change, and remove the old versions.
• Read the instruction block aloud and cut any line nobody can explain.
• Check that no rule appears in both the instructions and a knowledge file.
• Archive projects whose content type no longer recurs, so they stop cluttering the list.

Your Claude project knowledge base is a set of files you curate, so prune it like one. Archiving is safe, too: Anthropic notes that an archived project keeps its knowledge, members and permissions, and you can still open its conversations.

One project per content type

The setup that works at scale is one project per content type, not one per client, campaign or mood. The reason is context bleed: a project trained on newsletter work starts shaping blog drafts like newsletters, and a fresh project with the right instructions is cheaper than fighting the drift.

Splitting by content type also keeps each knowledge base small, which keeps Claude's reads relevant, and it's the main reason Claude Projects for content teams hold up better than one giant shared project. A newsletter project carrying SEO briefs is a project carrying noise. It's the same trade-off that decides how to scale content marketing without losing quality: more output only works when each stream keeps its own standards.

Memory adds a second reason to split. Anthropic's help center says each project has its own memory, and that a project also uses your account memory from chats outside projects unless you turn off "Use account memory" on the project's page. If you write for several clients, turn that setting off in each client project so one client's details never surface in another's drafts.

The cost of the split is maintenance: more projects means more instruction blocks to keep current. A project earns its existence when its content type recurs at least monthly with stable context. Rarer work, or anything whose sources decay weekly, lives in a plain chat with a pasted brief. For those one-off chats, a shared AI prompt library keeps the good briefs from getting lost.

A starter set for a content team

If you're setting up Claude Projects for content teams from nothing, the four projects below cover most of a team's recurring work. Each has its own instructions and knowledge, and each hands its output to the next.

The repurposing project is programmatic content at a small scale: one template per channel, a real source, many outputs. The reporting project needs a fresh data export uploaded at the start of each run, because a project only knows what you last gave it.

Start with the drafting project for your most frequent content type. It returns the most time, and its corrections teach you how to write instructions for the other three.

Sharing a project with your team

On Team and Enterprise plans, a shared project means every writer works from one set of instructions instead of five private copies. You can share by name or email, add people in bulk from a list of addresses, or make the project visible to your whole organization, and Anthropic's help center lists two permission levels.

"Can view" members see the knowledge and instructions and can chat inside the project, but can't change it. "Can edit" members can change the instructions and knowledge, manage members and contribute to the project. Give "Can edit" to the one person who owns the instructions, and "Can view" to everyone who just uses them.

That split matters more than it looks. Shared projects are where the handoffs in a content marketing pipeline stop leaking, because the researcher, the writer and the editor all work from the same rules. Five editors with "Can edit" will rewrite the block five ways in a month.

When a project is the wrong tool

Three situations belong outside projects, and knowing them settles most Claude projects vs chats decisions. One-off research is the first: a project's value is accumulated context, and a one-off has none, so a plain chat with a good prompt beats a project you'll never revisit.

Anything that depends on fresh sources is the second. A project carries the knowledge you gave it, and that ages. Work that depends on today's search results or this week's news belongs in a chat with web search, or in a scheduled job. The alerting side of that, deciding which ranking or indexing drop deserves a person's attention, is what automated SEO monitoring handles, and no project replaces it.

Recurring work that should run without you is the third. Projects wait for you to open a chat, while Cowork's scheduled tasks run in the cloud on a cadence, and Cowork projects can hold those tasks. The table sorts common content jobs into the container that suits each one.

The quick test: if the work recurs, the context is stable and one person owns the rules, a project pays. Miss any of the three and the project becomes a folder you maintain for nothing.

The new Projects, in beta

Anthropic is rolling out a new version of projects in stages, starting with Claude Code. In the new version a project is one conversation: you say what you need, Claude breaks the work into parallel threads that run in the cloud, and the threads keep going when you close your laptop. Every thread starts with the project's files, instructions and memory, and a Library collects the files you add and the files Claude produces.

Access is limited for now. Anthropic's help center says the beta is open to select Pro, Max and Ultra subscribers who use Claude Code, with a waitlist at claude.com/form/projects, and that chat, Cowork, Team and Enterprise will follow. When an existing project moves to the new version, its instructions, files, full chat history and scheduled tasks come with it, and nothing is deleted.

For content teams, that means the instruction block you write today carries over. Running several threads at once does use your plan's limits faster, so start with one thread per job until you know what your plan absorbs.

The Claude Code version of this setup

Once you know how to use Claude Projects, the file-based version is a short step. If you'd rather run this from files you own than from a claude.ai project, my free course covers the Claude Code version: six modules, 27 lessons, no signup. You finish with six files you keep, including a CLAUDE.md built from your own answers and an anti-slop kill list, which play the same roles as a project's instruction block and knowledge files.

If you want the whole content system owned rather than built yourself, that's what I do at busyless as a fractional Head of Content. Book a call, and if a good instruction block and an afternoon are all you need, I'll say so.