Claude Projects for Marketing: Repeatable AI Workflows
Stop re-briefing the AI in every chat. The four Claude Projects I'd build for a marketing team, what goes in each, and where the format hits its limits.
Most marketing AI use dies the same quiet death: every chat starts from zero, so every chat starts with five minutes of re-explaining the product, the audience and the voice, until one day nobody bothers and the drafts go generic. The fix is standing context, and I’ve written about building it in ChatGPT with system prompts. Claude’s answer to the same problem is Projects, and for marketing work I think it’s the cleaner design, because it separates the two things a working setup needs: instructions and knowledge.
A Project, per Anthropic’s announcement, is a workspace that groups chats around a curated knowledge base - documents Claude reads in every conversation inside the Project - plus custom instructions that tailor how it responds there. The knowledge base holds a 200K token context window, which Anthropic describes as the equivalent of a 500-page book, and paid plans stretch beyond that with a retrieval mode that searches your uploads instead of loading them whole, expanding capacity by up to 10x. Projects are available on every plan now, including free accounts, where you get up to five - which happens to be roughly how many a marketing operation needs.
Instructions versus knowledge, and why the split matters
The design decision that makes Projects work is that rules and reference material live in different places, and I’d respect the split deliberately. Project instructions are the standing rules: who Claude is acting as, your voice constraints, your claims policy. Project knowledge is the material those rules operate on: the brand guide, the positioning doc, past examples, product facts.
The practical difference shows up when things change. Product pricing changes, you swap one file in knowledge. Your voice rules don’t change, and they stay untouched in instructions. In a single mega-prompt - the ChatGPT custom instructions approach - every update means re-editing one fragile blob. In a Project, maintenance is file management, which is a thing a marketing team already knows how to do.
One rule I’d put in every marketing Project’s instructions regardless of purpose, copied from my system prompt template: never invent statistics, quotes, reviews or product claims - write [NEEDS SOURCE] instead. Standing context makes the model more convincing, and more convincing includes more convincingly wrong.
Project one: the copy desk
The first Project I’d build is the everyday drafting workspace. Knowledge: the brand voice guide, three to five examples of your best published copy, the product fact sheet, the banned-claims list. Instructions: the role, the voice rules as checkable constraints, and the output habits you want by default - variants labeled, character limits respected, questions asked when context is missing.
This is the brand voice setup made permanent, and the example files are doing most of the work - examples steer a model’s voice better than any description of the voice. The payoff is that “write the launch email” becomes a one-line request instead of a briefing session, in every chat, for everyone using the Project.
Project two: the content pipeline
The second Project handles long-form content, and its knowledge base is different enough to justify the separation: your content strategy doc, keyword and topic plans, the quarterly calendar and its topic bench, briefs, outlines of what’s already published so drafts link to it instead of repeating it. Instructions: your brief-first workflow rules - the model drafts only from an approved brief, states what it couldn’t verify, and never pads a thin outline into a long article.
Keeping drafting separate from the copy desk sounds bureaucratic and isn’t. The copy desk optimizes for speed on short assets. The pipeline optimizes for structure on long ones. My reason for the split is practical: in a shared workspace the instructions fight each other - the long-form rules make the short copy stiff, and the speed habits make the articles thin.
Project three: the research room
Third: competitive and market research, with knowledge holding competitor positioning pages you’ve saved, win/loss notes, customer interview transcripts and reviews. Instructions borrow from my competitor analysis prompts: every claim tagged to its source document, gaps stated as gaps, no speculation dressed as intelligence.
Transcripts are where the 200K window earns its keep, and where bigger archives trigger the retrieval mode on paid plans - Claude searches the project knowledge and pulls the relevant passages instead of reading everything at once. The practical effect for a marketer: “what do churned customers say about onboarding, with quotes” becomes a query against your own evidence rather than a memory test for a model.
Project four: the reporting narrator
The last Project turns numbers you paste into narrative your stakeholders read. Knowledge: metric definitions, goals, last quarter’s report as a format example. Instructions carry the one rule that makes AI-assisted reporting safe, the same one my measurement guides build on: numbers in, narrative out. The model may summarize, compare and flag, and may never supply a figure that wasn’t pasted in.
This is the Project I’d build last but use most reliably, because reporting is weekly, formatted, and exactly the kind of repeated-structure work standing context is for.
Sharing, and the quiet win for team consistency
On Team and Enterprise plans, Projects can be shared with view or edit permissions per member. I’d read that as more than a collaboration feature: a shared Project is enforced consistency. The junior marketer and the senior one draft against the same voice files, the same claims rules, the same examples. The alternative - everyone maintaining a private prompt collection of varying quality - is how a team ships four different brand voices while everyone follows “the AI guidelines.”
Ownership note that matters more than it looks: whoever manages the Project effectively owns the team’s marketing brain. Treat the knowledge files like the brand assets they are, with a named owner and a review date, or the Project rots the way every unowned wiki does.
Where Projects stop
I’d be dishonest to pitch Projects as the whole answer, so the limits, plainly. A Project can’t act - it holds context for conversations, and multi-step automated workflows need agent skills or tooling beyond the chat window. Knowledge doesn’t update itself - a stale positioning doc produces confidently stale output, which is worse than no context because it looks right. And standing context is not standing judgment: everything that leaves a Project is still a draft with the same verification burden as any AI output. The Project remembers your business. Whether the work is true remains your job.
FAQ
Are Claude Projects free? Yes - Projects are available to all users including free accounts, with free users capped at five projects. The retrieval mode for oversized knowledge bases and project sharing need a paid plan.
How much can project knowledge hold? The base is a 200K token context window, about a 500-page book. On paid plans, retrieval mode expands capacity by up to 10x by searching your files rather than loading them all.
How is this different from ChatGPT custom instructions or GPTs? The concepts rhyme, and my advice is the same on both platforms: build the standing context wherever your team already works. What I like about Claude’s version for marketing specifically is the clean split between instructions and a file-based knowledge base you maintain like a folder, plus per-project isolation so the copy desk’s rules never leak into the research room.
What should go into a Project first? The same thing I’d put in any standing setup first: your real voice examples and your claims rules. Context that prevents bad output beats context that decorates good output, and everything else can be added the week a task needs it.
This post is part of the prompts and workflows library. The standing-context idea in ChatGPT form is system prompts for marketing, the voice files come from the brand voice setup, and the wider practice starts at the complete guide to AI marketing.