Updating Your Prompts

Prompts control how your agents interpret questions and generate responses. Discover allows you to update the prompts within your agents at any time to refine behaviour, improve output quality, or adapt to new use cases.

---

How to Update a Prompt

1. Navigate to Agent in the left sidebar.

2. Click on the agent you want to modify.

3. In the agent’s workflow canvas, click on the component whose prompt you want to update (e.g., Generate, Rewrite Question, Coordinator).

4. The component’s configuration panel opens on the right side.

5. Find the Prompt or System prompt field and make your changes.

6. Click Save to apply the update.

You can test your changes immediately using the Run button at the top of the canvas.

---

Using Variables in Prompts

Prompts support dynamic variables using the {variable_name} syntax. These inject context from upstream components into your prompt at runtime.

To insert a variable:

• Type / inside the prompt field to see a list of available keys, or
• Type { followed by the variable name directly.

Common variables include:

• {knowledge} — chunks retrieved from your Knowledge Base
• {task_input} — the user’s original question
• {system_prompt} — a reference to the agent’s top-level instructions (used in orchestration prompts)

Ensure that any variable you reference in a prompt is produced by an upstream component in the workflow — otherwise it will be empty at runtime.

---

Do Not Modify the System Prompt

Do not modify the System Prompt

Do not change the System Prompt of an agent unless you fully understand the agent’s internal workflow.

The System Prompt governs the agent’s core behaviour and output format. It is carefully engineered to work with the rest of the agent’s components. Modifying it — even small changes — can degrade the agent’s performance, cause incorrect SQL generation, break multi-agent coordination, or produce unreliable responses.

If you want to customise how your agent behaves, use the appropriate fields instead:

Goal	Where to make changes
Change the agent’s tone or focus	Edit the Domain Knowledge field in the agent settings
Improve retrieval quality	Adjust descriptions in the Schema Manager or Knowledge Base
Add business context	Use the Domain Knowledge section in your Database Agent
Guide a Generate component	Edit the prompt in that specific Generate component
Add few-shot examples	Add them to the agent’s Domain Knowledge or KB

---

Prompt Fields by Agent Type

Different agent types expose different prompt fields:

Database Agent

• Domain Knowledge: Free-text field for business rules and context (e.g., “revenue = quantity × unit_price”). This is the safest place to influence SQL generation behaviour.

Coordinator Agent

Coordinator Agents expose multiple orchestration prompts. Each controls a specific stage:

Prompt	Purpose
Complexity Analysis Prompt	Determines whether a query is simple or complex
Agent Routing Prompt	Decides which sub-agent(s) to invoke
Result Synthesis Prompt	Combines results from multiple sub-agents

Each of these prompts uses specific required variables. Do not remove these variables — doing so will break agent orchestration.

Generate Component

The Generate component in custom agent workflows has a freely editable system prompt. This is the most flexible prompt in the system and safe to update, as it is designed for customisation.

---

Tips for Writing Good Prompts

• Be specific about the output format you expect (e.g., “Respond in bullet points”, “Return a JSON object with keys: summary, confidence”).
• Reference upstream variables to give the LLM the context it needs.
• Avoid vague instructions — the more precise the prompt, the more consistent the output.
• Test changes incrementally using Step Run on individual components before testing the full agent.