The Knowledge slot in a Claude Project is the difference between a tool that gives you average answers and infrastructure that knows your business. Here is what to put in it. And what to leave out.
Most marketing teams treat the Project Knowledge upload like a junk drawer. They upload the company deck, three random PDFs, and a screenshot of last year's plan, and then wonder why Claude keeps citing campaigns from two quarters ago. The Knowledge files are not a backup. They are the working memory of your Project. What you put in there decides what kind of answer you can get out.
The rule for what goes in: upload only files that change what Claude can say accurately about your business. Five to ten files. No more. Each one earning its slot. The seven files below cover almost every marketing team's needs.
A Project Knowledge file is good if removing it would make Claude's answers measurably worse. Most uploads fail this test. They are present because the team felt they should be present, not because they change the output.
Before you upload anything, ask: "What question would this file let Claude answer that it could not answer without it?" If you cannot name one, do not upload it.
The single highest-leverage file. Without it, every piece of output Claude gives you will sound like the average voice for your category. With it, every prompt downstream inherits your voice. This is the file that makes the team go "oh, Claude actually understands us." More on extraction in the brand voice post and the free Voice & Tone Extractor.
One page on exactly who you sell to. Job titles, company sizes, industries, the actual problems they describe in their own words, what they have tried before, what they would lose by not solving this. Not a buyer persona PDF with stock photos. A working operator's understanding of the person Claude is helping you talk to.
The actual text of your five best pieces from the last quarter. Not links. Not screenshots. The raw text. This gives Claude examples of what "good" looks like for your team specifically, so it can mirror the patterns when you ask for new content. The criteria for "best" is up to you: most conversions, most reach, most replies, whatever your team uses.
The one-page version of how you describe what you do, who it is for, and what you replace. Should answer: what category are you in, what do you do differently, who specifically benefits, what is the alternative. If your team has not written this down, do it before you upload, not after. Claude cannot extract a position from a logo and a tagline.
The honest version, including what underperformed. Numbers, what happened, what you learned. This file lets Claude reference real performance when planning the next thing, instead of inferring from category averages. Replace it every quarter. Stale readouts cause stale recommendations.
The output of running the free Competitive Landscape Teardown on your category, or your own equivalent. Three to seven competitors covered with positioning, pricing, ICP, and exploitable weaknesses. The point is that when you ask Claude for a campaign idea, it can cite which competitor weakness you are exploiting, not just guess that "differentiation matters."
What you measure, exactly how it is calculated, and what counts as good. Every team has a slightly different "qualified lead" definition, a different "engaged" threshold, a different attribution window. This file stops Claude from giving you advice based on the average definition, which is almost never your definition. Short, factual, lives forever.
The Knowledge files are not "upload once and forget." Three of the seven (performance readout, top 5 content, competitive teardown) need to be replaced every quarter. The ICP changes once a quarter or so. The positioning doc gets a check twice a year. The KPI definitions only change when the team changes how it counts.
The full quarterly maintenance ritual (system prompt, promoted prompts, KPI snapshot, do-not-write file, reference docs) is in the quarterly Project rebuild post. The 90 minutes once every three months is the difference between a Project that compounds in value and one that rots.
This post is what to put in the Knowledge slot once the Project exists. The setup walkthrough (the four-section system prompt, the multi-brand pattern, the stale-Project trap) is the prerequisite. And before you build any of this, the decision tree on Skill vs Project vs plain prompt helps you decide what belongs in a Project in the first place.
See the Project setup walkthrough