Claude chief of staff prompt: the AI setup for a biotech CEO
You can write a prompt that makes Claude behave like a chief of staff for one conversation. Getting it to behave that way next Tuesday is a different piece of writing, and this is what goes in it.
Search for a chief of staff prompt and you will find long paragraphs that begin "You are an experienced chief of staff to a busy executive." They work, for about twenty minutes. The reason they stop working is not the wording. It is that a prompt is a thing you paste into a conversation, and a conversation ends. Next week you paste it again, into a session that has never heard of your programs, your partners, or what you decided last Thursday, and you spend the first ten minutes re-explaining your company to a machine that will forget again.
The thing worth writing is not a prompt. It is a standing instruction that sits in a file, next to the records, and is read at the start of every session automatically. It is still just text you wrote in plain English. What changes is that it persists, and that it can refer to files rather than trying to contain everything.
What a prompt buys, and where it stops
| A pasted prompt | A standing instruction file | |
|---|---|---|
| Lifetime | One conversation | Every session, until you edit it |
| What it can hold | Whatever you retype | Pointers to the records themselves |
| Where the output goes | Back into the chat | Into the file it belongs in |
| What happens when you correct it | Correction lost at the end | The correction is written down and holds |
That last row is the one people underestimate. Most of the value in this setup comes from accumulating corrections. You tell it once that a message from the tox vendor goes to the program file rather than the vendor file, and that has to still be true in November. In a chat it is not.
The five things the instruction has to say
Keep it short. A page that tries to anticipate everything gets ignored in the parts that matter. Five sections is enough.
- Where things live. One folder per program, deal or partner, named the way you say them out loud. Say which file holds the current status, because that is the one it must never duplicate.
- What it may do without asking. File an arriving document, update a date the source material moved, close an item that was resolved, open a new one, write a draft. Everything reversible.
- What it must never do. Send anything. Delete anything. State something as settled without a source. Put a number in a summary that is not in a file.
- How to route what arrives. The rules below. This is the section that makes it yours rather than generic.
- How to answer you. First sentence is the answer. Cite the file. Flag what it could not verify instead of smoothing over it.
The one line worth stealing. "If you cannot point to the file that says it, say you could not confirm it." Almost every bad outcome in this pattern comes from a confident answer built on something that quietly went out of date, and one sentence of instruction changes the default from smooth to honest.
The routing rules that make it a biotech setup
This is the section that cannot be copied from a generic template, because the routing is the domain. A drug development company's inbound is dominated by organizations you do not control, and each of them needs a different destination and a different reflex.
- A message from a contract research organization about a study goes to that program, not to a vendor folder. If it carries a date, the date moves in the timeline and anything downstream of it moves too.
- Anything from a partner on a co-developed asset goes to the alliance file, and if it references a clause, quote the clause rather than paraphrasing it.
- A diligence request from a potential partner opens an item with the requested-by date on it, even when the sender did not give one, and especially when the data room does not have the answer yet.
- An invoice or change order from a vendor is matched against what was agreed, and a mismatch is flagged the day it lands rather than at the quarter close.
- Anything containing patient-level information stops. It is not filed, not summarized, not quoted. That is a hard boundary and it belongs in the instruction, in those words.
Write four or five of those in your own vocabulary and the setup stops being a persona and starts being a system. It is also the moment the thing becomes worth keeping, because the routing rules are the part that took you years to learn and thirty seconds to write down.
From an instruction to something that persists
Once the instruction is stable, the next two moves are packaging it so it fires on its own rather than because you remembered, and delegating pieces of it to sub-processes with their own context. Those are the subjects of the Claude chief of staff skill and the Claude chief of staff agent, and they are genuinely different things rather than three names for the same one.
The platform question underneath all of it, which surface of Claude you are actually writing this for and what each one can enforce, is Claude as an AI chief of staff, set up in a biotech. The role framing, if you are still deciding whether you want this at all, is an AI chief of staff for a biotech founder, and the concept underneath it is what an AI company brain is.
Routing rules that still hold in November
A standing instruction sitting beside the records, months of your own corrections still in force: the study message routed to the program rather than a vendor folder, the clause quoted instead of paraphrased, the change order matched against what was agreed, patient-level material stopped at the door. The demo is built on exactly that file.
See the demo