Prompt & output
Every modelling stage — Capabilities, Business Areas, Entities, Behaviour, Automations, Roles, Workflows, Agents — is driven by a call to the AI. The Prompt & output panel lets you see exactly what Kiln sends, tune it for a single session, and inspect what came back — so you can improve your results instead of guessing.
It's on‑demand: open it from the Prompt & output button in the stage's action bar, next to Generate and Second opinion. It opens as a drawer on the right and leaves the canvas in place.
1. See the prompts
The panel shows the two prompts behind the stage, clearly labelled:
- Generation — the instruction Kiln uses when you press Generate with LLM.
- Review — the instruction the Second opinion uses to critique the layer.
These are the real prompts, shown verbatim. Reading them is often enough on its own: it tells you what the model was asked to do, which explains why a result looks the way it does.
2. Tune a prompt for this session
Each prompt sits in an editable box. Change it and the next Generate (or Second opinion) on that stage uses your version instead of the default.
The edit is session only — it is not saved. The stored prompt is never changed, nothing is written to your project, and reloading the page restores the default. A Modified marker appears while your version differs from the default, and Reset puts it back. Use it to experiment: nudge the wording, run the stage, and see whether the result improves.
Because the edit is session‑only, it's a scratchpad, not a setting. If you find a wording you want to keep, note it down — it won't survive a reload.
3. Compare against the last output
Under Last output, the panel keeps the most recent result for that stage — both the generation and the review, each overwritten on the next run. Alongside it you'll see which model and engine produced it, when, and whether a tuned prompt was in effect.
That closes the loop: view the prompt → tune it → re‑run → compare the new output to the last one. It's how you tell whether a change to the prompt actually made things better.
The last output travels with your model: it's kept when you export model.json and comes back when you import it, so you can pick up a comparison later. It's an inspection record only — it never becomes part of your model, and the code export ignores it entirely.