Most Claude Projects rot in production. The 90-minute quarterly refresh that keeps them from becoming stale brand-shaped noise generators.
A Claude Project set up once drifts in three months. The system prompt that captured your voice in January is roughly your voice in April. The KPI snapshot you uploaded then is now three quarters of revenue out of date. The prompts you actually used 50 times last quarter are not in Project memory at all. The model is fine. The Project is the problem.
I rebuild mine every quarter. Ninety minutes, end of quarter, before the new quarter starts. The output is not a new Project. It is the same Project sharper, with the stale parts cut out and the patterns I actually used promoted in.
The cost of not doing this is the slop tax in slow motion. Every output the Project produces inherits whatever drift accumulated since the last refresh. By month nine that drift is real, and somebody on the team will mistake it for the model getting worse. The model did not get worse. Your Project did.
The drift is invisible until it is not. Your team starts editing Claude outputs more heavily than they did before. The brand voice in drafts feels off, but nobody can name how. The KPIs in performance write-ups are last year's KPIs, because that is what the Project still has.
None of these get flagged as "the Project needs refreshing." They get flagged as "Claude is not as useful as it was." It is the same complaint that arrives when a sharpened knife has not been sharpened in nine months. The blade did not change. The use of it did.
The list is short on purpose. Quarterly refresh is a 90-minute job, not a full rebuild.
What has changed about the voice, the audience, the current quarter's focus, the metrics that matter, the things you have decided not to do.
The single biggest update most quarters: the non-goals section. Marketing teams generate non-goals constantly. Campaigns killed. Positions abandoned. Audiences deprioritised. Almost nobody updates the Project to reflect them. So the Project keeps offering to help with the thing the team agreed three months ago to stop doing.
Open your chat history for the quarter. Find the five to ten prompts you actually ran more than three times. Promote them into Project memory as named patterns.
Stop typing the same prompt structure ten times a week. The Project should know your repeated workflow. If you find yourself writing "I want a brief that..." for the fourth time, that is a prompt that belongs in Project memory.
The flip side matters too: the prompts you thought you would use, that the Project memory still references, that you have not actually run all quarter. Cut them. A Project optimised for the workflow you imagined is worse than a Project optimised for the workflow you have.
The numbers in the Project should be the current quarter's targets and the previous quarter's actuals. Not last year's. Not the snapshot you uploaded when you set the Project up.
Why this matters: when you ask the Project for a campaign brief and the success criteria are wrong by 40%, the brief is wrong by 40%. Stale KPIs produce strategy that aimed at the wrong floor.
The list of things you have decided not to do, not to say, not to recommend. Campaign types that did not work. Positions you used to test and abandoned. Words the brand decided are off-limits.
Marketers accumulate these constantly and almost never write them down anywhere the Project can see. So the Project keeps offering them up, because as far as it knows, they are fine. A quarterly pass to capture this turns "we tried that and it did not work" tribal knowledge into actual Project constraint.
Look at what is uploaded to the Project. Which docs are now stale and should come out. Which new docs (brand voice updates, new positioning, last quarter's campaign post-mortems) should go in.
The temptation is to add. The discipline is to subtract. A Project with 40 reference docs from various points in time produces output that averages across all of them, which means averaged voice, averaged positioning, averaged advice. Cut to the ten that actually represent who you are right now.
The test is downstream. After a quarterly refresh, two things should change in measurable ways.
The amount of editing on Claude outputs drops. If you were rewriting 60% of every draft and now you are rewriting 20%, the refresh worked. If you are rewriting the same 60%, the refresh missed.
The team's questions about Claude shift. "Why is Claude suggesting X" or "Claude keeps using this old framing" should disappear after a refresh that addressed those drifts. If those questions stay, you missed updating the thing that drove them.
If neither of those changes, the refresh was probably surface-level. Go deeper.
Monthly is overkill. Most marketing programs do not change enough in 30 days to warrant rebuilding the Project around new state. You would just be re-touching the same things.
Yearly is too long. Three quarters of accumulated drift is the difference between a Project that helps and one that confuses. By month nine, the Project's brand voice is a costume of who you were when you set it up.
Quarterly maps to how marketing teams already plan. The end of every quarter is already a moment when KPIs reset, campaigns get killed or extended, and the team makes calls about what is next. Folding the Project refresh into that existing rhythm costs almost nothing in marginal time. Doing it as its own thing on a random Tuesday rarely happens.
A Claude Project compounds in value if you maintain it, and rots if you do not. Most teams do the first 10% and skip the maintenance, then blame the model. The model is fine. Maintain the Project.
The setup walkthrough (four-section system prompt, voice + KPIs + tools + non-goals) is the prerequisite. This post is what to do once it has been live a quarter and started to drift.
See the Project setup walkthrough →