Prompt templates

Each AI-generating pipeline step runs from a prompt template — the system prompt, context, task instruction, and a typed output contract that the model’s response must satisfy. Templates are per-account, so you can tune voice and behaviour without affecting other accounts.

🎬 Edit a prompt template

Prefer to read? Open the step-by-step transcript

1. Open Settings → Prompt templates.
2. Pick the step’s template to edit.
3. Adjust the instruction text; the typed field contract defines the required output fields.
4. Save — new productions use the updated template; in-flight ones keep the version they started with.

What a template contains

• System prompt — the model’s role and ground rules.
• Context tokens — placeholders (e.g. the brief, scene, persona) filled in at runtime.
• Task + output instruction — what to produce and in what shape.
• Typed field contract — a schema the model output must validate against. If the output doesn’t match, the step fails with a contract violation rather than passing bad data downstream.

Contract changes are sensitive

Tightening a typed field contract can make previously-valid outputs fail. Roll out contract changes carefully (staged across accounts) and after the supporting code is in place — otherwise productions hit contract-violation failures. For most tuning you only need to edit the instruction text, not the contract.

Editing safely

1. Change the instruction, not the contract, for tone/behaviour tweaks — lowest risk.

2. Test on one account first. Run a production and check the step output (the production detail page shows each step’s result).

3. Roll out gradually if you’re applying a template change across many accounts — staged, not all at once.

Note

Each production captures the template version it ran with, so editing a template never rewrites the history of completed productions.