Instructions that scale
Outcomes
Skills you switch on
Instructions are the context source you shape most directly, and the platform organizes them into capabilities: discrete skills you enable for an AI Employee, like capture a lead, book an appointment, or answer from knowledge. Each capability is a set of instructions, sometimes with a tool attached, and an employee can carry several at once. Here is where they sit in the whole:
The employee reads each situation and applies the capability that fits: a question about hours draws on knowledge, interest in a service activates lead capture, a booking request activates appointment booking. You choose the skills; the employee chooses the moment.
Ground rules and the moment
If you have used any AI chat tool, you have felt the difference between two layers of communication. There are the ground rules you would set once ("you are a data analyst; keep answers concise and plain") and there is the request of the moment ("summarize this report"). The ground rules stay constant; the requests keep changing.
An AI Employee has exactly this structure. Capabilities are the ground rules, written once and present in every interaction. The customer's message is the request of the moment. This is why instructions are worth real craft: a conversational request is used once, but a capability instruction runs on every conversation, every day, for as long as the employee works. Write it well once and it scales.
Behavior or facts
A two-question test sorts anything a business wants its employee to handle:
- Should this govern behavior, or apply every time? Instructions. "Always collect a name and phone number before booking" is behavior.
- Is this a fact to look up when relevant? Knowledge. The list of services and prices is facts.
Run the test and the pieces land where they work: rules where they are always present, facts where they are retrieved when needed.
Short beats long
Everything an AI Employee is told, across all of its capabilities, lands in one working memory. Instructions do not queue up politely and wait their turn; they are all present at once, together with the retrieved knowledge and the conversation. Two long capabilities with overlapping instructions compete with each other, and the answers show it. Partners in the field have a phrase for an employee configured this way: it gets its wires crossed.
The practice that follows is simple and reliable: keep instructions short, specific, and action-oriented. Every sentence in a capability should change what the employee does. If a sentence repeats another capability, restates the obvious, or narrates background, cut it. A tight capability is easier for the model to follow and easier for the next person to maintain.
Brief it like a new team member
The instructions themselves have a craft, and it is the same craft as briefing a new hire on a skill. A strong capability instruction answers three questions, in order:
- When do you act? The trigger: "when a visitor expresses interest in a service."
- What must be true first? The requirements: "collect a name, and a phone number or email, before proceeding."
- What does done look like? The outcome: "confirm the details back, and let them know someone will follow up within one business day."
Structure helps too: short lines and numbered steps outperform paragraphs, for models and for the humans who maintain them.
Pick one thing a client's receptionist does every day and write it as a capability brief in three sentences: when to act, what must be true first, and what done looks like. That is instruction-writing, and you just did it.
Knowledge Check
Three quick questions on capabilities, placement, and instruction craft.