The situation
Seven weeks into my CMO search, Claude had stopped being a chat window. It runs a scheduled morning brief, a five o'clock search summary, a weekly rebuild of my search board, the wildfire tracker on this site, and a few short-lived monitors. Six data sources feed it. Four custom skills turn one-word commands into full runbooks. Nothing that grows that fast has been audited. So I asked the system to audit itself.
The prompt
One paragraph, typed into the Claude desktop app:
Do a comprehensive review of how I'm using you. What am I doing well? Is it cutting edge for someone in my position? Where should I be using you that I'm not? Where am I trying to work with you but using an approach that's suboptimal? Give any other insights that you think would be helpful. Deliver your findings with some good looking visuals.
No rubric, no structure, no example of the output I wanted. An open question and access to everything.
What it did with it
Before writing a word, it read the evidence: twenty memory files, eleven scheduled tasks, sixteen published artifacts, and every skill and connector. Then it published a visual report.
Its headline was that I do not use Claude as a chat tool; I have built an operations layer. It scored eight capabilities from zero to five, noting that the scores are its read and not a platform metric:
- Scheduled automation, command skills, memory architecture — 5.
- Live surfaces, connected data — 4.
- Verification loops — 3.
- Meetings and calls — 1. A transcription connector is wired in and sits outside every workflow, while the search runs on calls.
- Publishing in my own byline — 1. My voice is profiled and nothing is published with it.
The more useful half was mechanics. The morning brief had minted twelve of my sixteen published artifacts as separate copies instead of updating one page. The file holding live search status had been rewritten five times in a month against a 32KB memory cap, once with 136 bytes to spare. I could not have seen either from inside a chat. One lived in a gallery I never scroll; the other in the logs of runs that happen while I sleep.
What we fixed in the same session
I asked for the top two fixes.
The morning brief now has one pinned page, republished in place every weekday. The URL never changes.
The search status moved out of memory into a JSON state file in Google Drive. The memory file became a short pointer carrying the read-and-write protocol, so the size ceiling stops mattering. Three scheduled task prompts and two interactive skills were rerouted to it. One wrinkle surfaced in testing: the Drive connector cannot read raw JSON back, so the state lives inside a Google Doc, refreshed by writing a new copy and trashing the old. That evening's scheduled summary was the first live run against it.
What I would tell another CMO
Ask the tool to grade you, and give it the evidence rather than your own account of it. The compliment in the first paragraph was pleasant. The ranked list with a specific fix beside each item was the part worth having, and two of the fixes were done before I closed the laptop.