CHouseUI
Using CHouse UIAI agents

Prompt templates

Feature variables, conditional sections and skill includes in system prompts and task templates, and how the preview checks them.

In the app
AI Governance › Assistant › AI Governance › Prompt/ai/assistant
Needs any of
ai_agents:viewai_agents:manage

An agent's system prompt and task template are templates. They are deliberately logic-free: three forms are interpreted, and every other character — including other {{…}} text such as {{slot_start}} — is kept as written.

FormEffect
{{ctx.name}}Insert the feature's variable name
{{#ctx.flag}} … {{/ctx.flag}}Keep the text between the tags only when flag is set
{{skill:name}}Insert the skill's SKILL.md (trimmed)
{{skill:name/file.md}}Insert one file of the skill (trimmed)

Variables

A feature hands its agent a fixed set of variables: the query being optimized, the node under investigation, the evidence it gathered, and so on. Each feature's variables, with their types and values, are listed in the AI feature catalog. A few examples:

FeatureVariables
Optimize queryctx.query, ctx.additionalPrompt
Diagnose a server errorctx.node.id, ctx.node.name, ctx.error.code, ctx.error.name, ctx.error.message
Fleet Doctor scanctx.hours, ctx.nodeCount, ctx.overview (JSON), ctx.needsPlaybook (boolean)
DataOps and observability featuresctx.evidence (JSON)

Rules:

  • Only feature agents (with a task template) get variables. A chat agent or a subagent that uses {{ctx.…}} is rejected.
  • A template may only use variables that every feature bound to the agent provides. The editor's palette shows exactly those.
  • A variable that isn't provided at run time is an error, not an empty string: the run fails closed rather than sending the model a broken prompt.
  • Values are inserted as plain text; JSON variables arrive as JSON text.

Sections

A section keeps its text when the variable is set: a non-empty string, a number other than zero, or true. Empty strings, 0, false and missing values drop the section. Sections can nest, and each must be closed by the same name.

The built-in SQL Optimizer uses one to add the user's extra instructions only when there are any:

text
Optimize this ClickHouse SQL query:

```sql
{{ctx.query}}
```
{{#ctx.additionalPrompt}}

Additional instructions from the user:
{{ctx.additionalPrompt}}{{/ctx.additionalPrompt}}

The Fleet Doctor pulls in the ClickHouse playbook only when the scan found a heavy query worth optimizing:

text
{{#ctx.needsPlaybook}}{{skill:clickhouse-playbook/reference.md}}{{/ctx.needsPlaybook}}

Skill includes

{{skill:…}} inlines a skill's text where the token stands. It is the same text a pinned skill adds, but at a place you choose — inside a section, before the rules, in the task template. The skill must exist and be enabled, and the file must exist in it, or the save is rejected; disabling a skill that a template includes shows up under Problems. The palette's Inline a skill… menu lists every enabled skill and file.

What the model receives

For a feature agent, the system prompt is assembled in this order:

  1. the rendered system prompt
  2. each pinned skill file, in the order the skills are listed
  3. the feature's output contract — the JSON schema the answer must match (not editable)
  4. the harness's prompt suffix

The rendered task template is the first user message. For a chat agent it is the same without the output contract and the task: the user's conversation follows the system prompt. Each subagent's prompt is rendered the same way with its own pinned skills and its own harness suffix.

The test console shows the exact system prompt and task message of a real run.

Preview and validation

Preview in the editor renders both templates against the bound feature (or the feature you opened the editor from) with placeholder values — strings show as «name», booleans as true so every section shows — and lists every problem the change would introduce. A feature agent that isn't bound yet has no variables to render with, so its preview reports its variables as unknown — try it in the test console instead, which picks the feature. The checks:

  • a variable the feature doesn't provide (Unknown template variable ctx.qurey)
  • a section that is never closed, or closed by another name
  • an unknown or disabled skill, or a file the skill doesn't have
  • {{ctx.…}} in a chat agent or subagent

The same checks run on save for the agent and for every feature it serves, directly or as a subagent.

Edit this page on GitHub ↗

Next →Harnesses