The trigger
An incident has been notified internally this morning. People are safe, risks are controlled, and if the incident may be notifiable the accountable person has already phoned the regulator; that duty never waits on drafting. Workers on that site need to hear what happened, today. This recipe chains the incident triage skill file into the safety comms skill file to get there.
The inputs
- The first-hour incident record, de-identified: roles, not names
- Site, and times of the incident and of becoming aware
- Injuries or exposures by role, and any plant or substances involved
- The one settled action for workers, if decided
- The contact who answers worker questions
- The accountable person who owns notifiability and sign-off
The steps
- Confirm with the accountable person that regulator notification has been handled first; drafting never delays that call.
- De-identify the incident report before it touches the model, replacing names with role tokens.
- Run the structuring prompt to lay the facts under each notifiable incident question, unknowns marked UNKNOWN.
- Hand the structured record to the accountable person for the notifiability decision, which the AI output leaves unanswered.
- Confirm the one action workers must take; if several controls are circulating, settle which leads.
- Run the drafting prompt to produce the alert and toolbox talk intro.
- Take both drafts to the HSR and the accountable person for sign-off, then release.
The prompts
Run the structuring prompt first; draft only after the human decisions are made.
The checks
- Confirm every fact appears in the source incident record; strike anything the model added.
- Check times and the site against the original notification, not the AI text.
- Confirm unknowns stayed marked UNKNOWN.
- Scan both drafts for names or identifying detail that survived de-identification.
- Confirm one clear action, no blame language, and no sentence implying a notifiability conclusion.
Governance controls
- The notifiability decision and the regulator notification belong to the accountable person; AI never decides or scores notifiability, and never delays the immediate call.
- Inputs are de-identified before drafting: roles not names, placeholder tokens throughout, the standard set in human-in-the-loop incident analysis.
- The alert is reviewed and signed off by the named accountable person, with HSR consultation, before it reaches workers; every output is a draft until then.
- File the signed-off alert with the notification record, kept for at least five years; the accountable person confirms jurisdictional requirements with the WHS regulator.
The output
A signed-off safety alert and toolbox talk intro, released the same day, plus a structured incident record on file.
Time it takes
About 25 minutes once the incident record exists. A first run takes longer while prompts are localised; a practised run drafts in minutes, with the rest spent on review gates.
TheAICommand. Intelligence, At Your Command.