Skip to main content
A task configuration describes one repeatable content job. This page walks through the editor top to bottom and builds a simple product description task. Open Task configurations and click Add config.

​
1. Choose an AI specialist

The specialist is the role the model plays, and the model it runs on. Pick one that matches the job — a content writer for marketing copy, an extractor for structured attributes. The specialist prompt describes who the AI is and how it behaves. It does not describe the output format, the field limits, or the quality criteria — the platform injects those from elsewhere in the configuration.
Resist the urge to write everything into this prompt. A requirement repeated in the prompt and in a quality rule will drift out of sync, and the rule is the one that is actually enforced. Where each requirement belongs

​
2. Add variants

A variant is a dimension the output varies across — language, brand, audience, persona. Each dimension has options, and each option carries fields describing it.
When you start a session you pick one option per dimension, and each combination produces its own output. Two languages and two audiences means four outputs per record.
Dimension or field? Add a new dimension only when the thing genuinely needs its own output per option. If it is extra information shaping the same output — tone notes, reading level, terminology — it belongs as a field on an existing option. Variants
Keep to 2–3 fields per option. Every field is a constraint the model has to hold, and if removing a field would not change the output, remove it.

​
3. Give the AI extra abilities

Under AI can also… you attach capabilities — looking up data in an index, calling an external MCP server, searching the web, or seeing the document’s images. Most first configurations need none of these. Add them when the job genuinely cannot be done from the document alone. Capabilities

​
4. Write the quality checks

This is the part that makes the job trustworthy. Three rule types are available: Start with a schema rule. Every configuration should have one, because it is the only thing that machine-enforces the shape of the output:
Route every black-and-white criterion to code. Models cannot reliably count characters or match strings; a schema rule does both in under ten milliseconds at zero token cost. Save AI evaluation for what actually requires language understanding. Quality checks
Keep to around five rules. Quality checks are for what must be verified every single time, not a wish list — more rules mean more prompt weight, more cost, and a twitchier gate.

​
5. Set the quality gate

The gate decides how the individual scores add up to accept or reject. Balanced is the right default and is described in plain language in the editor. Quality gates

​
6. Add special instructions (optional)

Free text for generation-only context — campaign notes, edge cases, one-off guidance.
Quality checks never see special instructions. Anything that must be verified does not belong here.

​
Save, preview, publish

Save Changes stores a draft. Preview runs it over a few documents so you can see real output. Publish makes the version live. Do them in that order, every time. Preview and review