Customising AI prompts
AI prompts are advanced and optional — the bundled default preset needs no configuration — but an organisation that wants different wording can pick a different preset or write its own instruction, one feature at a time.
What it is
Three features can be customised independently: AI image analysis, executive summary, and AI polish. For each, you can either select a ready-made preset (a complete, tested set of instructions for all three features together) or write a custom instruction for one specific feature that supersedes the preset just for that feature. Your survey data is always inserted server-side into fixed tokens the instruction text references (for example {{ROOM_NAME}}, {{FINDINGS}}, {{TEXT}}) — you edit only the surrounding instruction, never the mechanism that inserts real data into it. Leaving a box blank keeps that feature on the selected preset.
Before you start
Customising prompts needs organisation-admin access to Settings, one level lower than the bar for connecting or changing the AI provider connection itself.
How to do it
- Open Settings → AI and scroll to AI prompts (advanced — optional).
- Choose a preset from the options shown — each is summarised in a card so you can see what it’s for before selecting it.
- For any of the three features, type a custom instruction in its box to override the preset for that feature alone; leave it blank to keep using the preset.
- Click Reset to preset on any feature to discard a custom instruction and revert to the selected preset.
- Click Save AI prompts — this applies to your organisation only, on drafts, advisory only, exactly like every other AI feature.
How it integrates
A custom instruction cannot widen what a feature is allowed to produce: the result vocabulary a hazard’s photo-analysis feature can return stays fixed by that hazard’s own registered result-value list regardless of what the prompt’s instruction text says, and every AI call still runs through the same drafts-only, sealed-record-refusing path described in How AI is used, and its boundaries. What a custom prompt can change is tone, structure, and the emphasis of the instruction itself — for example, asking for shorter paragraphs, or a different register of language. The platform validates a saved prompt bundle before accepting it, rejecting anything missing a required token placeholder or missing a hazard’s own registered preset, so a malformed custom prompt fails at save time rather than silently breaking generation later.
Common problems
- I wrote a custom prompt and results didn’t change as expected. Confirm you edited the box for the right feature — image analysis, executive summary and AI polish are customised independently of each other.
- Save fails with a validation error. The bundle is checked for the token placeholders each feature substitutes into and for coverage of every registered hazard preset — a prompt missing one is rejected before it can silently break generation.
- I want to go back to the default. Click Reset to preset on each customised feature, or reselect the default preset itself.
Related