Claude chief of staff skill: AI for biotech operations
A skill is a capability with a name, a trigger and a fixed procedure. The point is that you stop remembering to invoke it, because it fires on the shape of what you asked.
People arrive at this phrase after the prompt stops being enough. A prompt is text you paste. A skill is a packaged procedure that sits alongside your setup and activates when the request matches what it is for, which means the good behavior happens whether or not you thought to ask for it. Anthropic lists Skills as available across Claude's surfaces, in chat, in Cowork and in Claude Code, so this is a platform capability rather than something bolted on.
The difference matters in practice more than it sounds. The failure mode of a prompt-based setup is not that the model is bad at the task. It is that on the Thursday when everything is on fire you do not paste anything, and that is exactly the week the record needed to keep itself.
The six that carry an operating week
| Skill | What sets it off | What it needs |
|---|---|---|
| Intake | A document or email arriving that belongs to a program, deal or agreement | A mail connector for the automatic version; works by hand without one |
| Meeting to record | A transcript or rough notes from a call | Nothing beyond the transcript |
| Sent-mail sweep | You sending something that makes a promise or asks for one | A mail connector |
| Follow-up sweep | A date passing on something you are owed | A scheduler and a machine that is awake |
| Person prep | A meeting appearing with someone you have history with | Calendar access for the automatic version |
| Morning brief | The clock | A scheduler; web access if you want the news layer |
The right-hand column is the honest part and it is usually missing from pages like this. Several of these are only as automatic as the connectors underneath them. Without a mail connector, intake is a thing you hand it rather than a thing that happens. Without a scheduler and a machine that is on, the overnight sweep is a command you run rather than a pass that runs, and "the moment it lands" means the next scheduled pass rather than the instant. All six also run for one operator: one set of files, one owner, one mail identity, so a team is served today by a seat each rather than by one shared brain, which has consequences covered in an AI coworker in Claude for a biotech team.
What the intake skill does in a drug development seat
Worth walking through one, because the generic description hides the value. Late on a Friday the partner on your co-developed program sends one message with three attachments and no explanation: a signed scan, a deck named for its version, and a numbered invoice. The intake skill opens each and puts it where it belongs rather than where it landed. The scan is the countersigned governance charter update, so it goes to the contract file for that collaboration. The deck is the draft pack for the September steering committee, so it becomes the current pack on the alliance record and opens exactly one item for you: check it against the agenda that locked on the fourteenth, before the meeting. The invoice is a cost-share charge, so it goes to the finance record against that same collaboration. Then it drafts your acknowledgment and stops.
Nothing in that chain is clever. All of it is work somebody currently does badly on a Friday afternoon, and the reason it is worth packaging as a skill rather than a prompt is that the chain has to run the same way every time or the record stops being trustworthy. The program-level view of this load is AI for program management in drug development.
The line the skills do not cross. Every one of them stops before anything irreversible. The reply is drafted and staged, not sent. Nothing is deleted. Nothing is signed or paid. The system takes the reversible step and hands you the decision, which is the only version of this that survives contact with a company that has any kind of review process. Reversible here means a daily snapshot you can roll back to, not a per-change undo button.
What sits below skills, in the place where a rule has to hold whether or not the model cooperates, is a different mechanism again, and it exists on only one of Claude's surfaces. That, and which surface to build on, is Claude as an AI chief of staff, set up in a biotech. What sits above them, where a job is handed to a separate process with its own context, is the Claude chief of staff agent. The instruction file all three read is the Claude chief of staff prompt.
If you have not yet decided whether to assemble this yourself or have it installed, the comparison is Claude Code for biotech vs off-the-shelf AI tools, and the concept underneath all of it is what an AI company brain is. Assembling the six from scratch is the long road. Installing the app is about thirty minutes of work, yourself or done with you. It is in limited early access, so getting a copy starts with a conversation.
Three attachments on a Friday night, filed three different ways
The countersigned charter goes to the contract file, the September pack lands on the alliance record with one item opened against the agenda that already locked, the cost-share invoice goes to finance, and the acknowledgment is drafted and left for you. All six triggers are running in the demo, with what each one needs underneath it said out loud.
See the demo