Skip to main content
A prompt is mostly fixed text, with gaps that are filled in each time the agent runs: the question someone asked, the name of the person who wrote, what an earlier step found. The gaps are written in double curly braces, {{ }}. ServFlow calls this dynamic content. This page explains how to read and write it. For every function, see the Dynamic content reference.

​
Reading a template

Text with gaps in it is called a template. This one is the User prompt of a Discord agent:
Everything outside the braces is sent as written. Each pair of braces is replaced when the agent runs: the first with the name of the person who wrote, the second with their message.

​
Where you can use it

In a System prompt and a User prompt, in a tool’s settings, in a context step’s settings, and in a guard’s condition. A field that accepts dynamic content shows Supports {{ }} templates under its label.

​
What you can read depends on how the agent was started

Three things to watch:
  • A Discord agent cannot use body. It has no web request to read.
  • body reads a web request only when the request is sent as JSON.
  • An agent that runs on a schedule has no message to read, so leave input out of its prompts.
The Available entry handlers lists every name under .discord, .telegram, and .github.

​
What you can read from any agent

  • What an earlier context step found. {{ .customer.name }} reads name from a step called customer. See Context fetching.
  • A stored secret. {{ secret "weather-api-key" }}. See Secrets.
  • Inside a tool, what the agent passed to it. {{ tool_param "question" }}. See Add a tool to an agent.

​
Small changes to a value

A function can change a value on its way in. Three common ones: The brackets in (body "count") mean “work this out first”.

​
If something is missing

A gap whose value is missing is left empty, and the prompt still goes out. Write prompts that read sensibly when that happens. A misspelt function name is different. In a guard, ServFlow refuses it when you save. In a prompt, the agent fails when it runs.

​
Dynamic content in the dashboard

Every field that accepts dynamic content has two small buttons in its corner. Insert a variable opens a menu of what the field can read, such as the results of context steps and your secrets, and writes the braces for you. Expand editor opens the field in a larger window. Typing inside braces suggests names as you go.
The Insert a variable menu open on a prompt, listing the results of two context steps
The expanded editor showing a User prompt that reads two context steps and the request body

Dynamic content reference

Every function, with what it takes and returns.

Available entry handlers

Every name a Discord, Telegram, or GitHub entry provides.

Guard rails

Conditions that decide whether a tool may run.

Context fetching

Steps whose results a prompt can read.