Set up NotebookLM for WHS practice

What this sets up

NotebookLM, which Google has renamed Gemini Notebook, works differently from a chat assistant: it is source-grounded, answering from the documents you load into a notebook rather than from an instructions field. This configuration uses that shape deliberately. You build one WHS reference notebook whose sources are your duties map, your templates and your codes of practice, plus a rules note that carries your working discipline, and because the rules live in a source, you restate them in each prompt's framing rather than relying on a persistent instruction. The output set is the WHS staple trio: incident summaries, hazard register entries and toolbox talk outlines. The method comes from the AI Setup module LM-S01; this page packages the WHS playbook for NotebookLM specifically.

What never goes in

Incident material is de-identified before it is uploaded or pasted, without exception: injured workers and witnesses appear as [WORKER_A], [WITNESS_1] and so on, and medical detail stays out entirely. Google's documentation adds these platform-specific reasons:

• Sharing a notebook shares its sources: "A viewer has read-only access to all the source documents and notes you shared with them in the shared notebook", and an editor "can view, add, or remove sources and notes in your shared notebook as well as share it further with other users". Chat View is not a privacy control: Google warns it "does not completely revoke a viewer's underlying access to the notebook contents". Owners and editors can also make a notebook public. Load nothing into a notebook you would not hand to everyone it could reach.
• On a consumer account, Google states notebook content "will not be used to directly train our foundational AI models, unless you choose to provide feedback" - and feedback is the trap, because it collects "the associated content - like your prompts, customizations, sources, uploads, and the outputs received". Do not use thumbs up or down in a notebook holding work material on a consumer account.
• On Google Workspace accounts the position is stronger: "Your uploads, queries and the model's responses in Gemini Notebook will not be reviewed by human reviewers even when you provide thumbs up or down feedback, and will not be used to train AI models." That carve-out belongs to Workspace and Workspace for Education accounts; the enterprise-grade protections attach to the editions Google lists, so confirm which edition your organisation runs.
• Feedback data, where collected, is "retained for up to 3 years", and Google's pages publish no general retention figure for uploaded sources and notebooks, deferring to the applicable privacy policy. Treat uploads as persistent until you delete them, and note that deleted notes cannot be recovered.
• Consumer notebook data shared with other Google services is used per the Google Privacy Policy "including for product improvement", so the training statement above does not cover everything that can happen to consumer data.

Sources, all verified 14 August 2026: Privacy and Terms of Use in Gemini Notebook, Use Gemini Notebook with a work or school Google account, Create a notebook, Share a public notebook, Gemini Notebook FAQ.

The instruction text

Create a notebook named "WHS reference", and add this as its first source, a note titled rules-note.md. Because NotebookLM grounds answers in sources, also open each working prompt by pointing at it, as the prompts below do.

Working rules for this WHS notebook.

1. All incident material in this notebook is de-identified.
   Injured workers and witnesses appear as [WORKER_A], [WITNESS_1]
   and so on. If identified details appear in any source or
   pasted text, the response must say so and stop.
2. Incident summaries follow incident-summary-template.md: facts
   first, sequence of events, contributing factors, actions.
   Separate observed fact from inference, and label every
   inference as one.
3. Hazard register entries use the exact field names in
   hazard-register-fields.md so they paste cleanly into the
   register.
4. Toolbox talk outlines follow toolbox-talk-format.md: one hazard
   theme, one real de-identified example, three questions for the
   crew, under ten minutes of content.
5. Never state that a duty has been discharged or breached. Frame
   legal characterisations as matters for assessment by a
   competent person.
6. Every draft ends with "For human review:" listing the points to
   verify against source records.

The files to attach

Load these as sources alongside the rules note. Each links to a published artefact to build it from:

• whs-duties-map.md: the duties and codes you work under, in plain language, jurisdiction noted. Start from the model WHS laws glossary entry and officer due diligence for AI work systems.
• incident-summary-template.md: facts, sequence, contributing factors, actions. The incident-to-safety-alert recipe shows the downstream use, downloadable as raw markdown.
• hazard-register-fields.md: exact field names and allowed values from your register, with the data discipline in AI for incident analysis and leading indicators.
• toolbox-talk-format.md: your talk structure, informed by the safety communications skill file.
• writing-style.md: your voice profile, built with the voice-profile interview in module LM-S01, Part 1.

The first three prompts

Apply rules-note.md. I am pasting a de-identified incident report.
Extract the timeline and the observed facts, then structure an
incident summary against incident-summary-template.md with facts
and inferences separated and every inference labelled. Do not
characterise any duty as discharged or breached. End with "For
human review:" listing the points to verify against source
records.

Apply rules-note.md. From the de-identified inspection notes I
paste, extract each hazard and map it to the exact field names in
hazard-register-fields.md, one entry per hazard, ready to paste
into the register. Flag any hazard where the notes do not support
a field value instead of guessing.

Apply rules-note.md. Using the de-identified incident example I
name, draft a toolbox talk outline against toolbox-talk-format.md:
one hazard theme, the example, three discussion questions for the
crew, under ten minutes of content, in the voice set by
writing-style.md.

Governance guardrails

• De-identify before upload, always: a notebook is a document store, and sharing it shares the documents. Keep the notebook unshared by default, grant viewer access only when needed, and never grant editor access casually, because editors can reshare and can generate public links.
• Risk ratings, control decisions and notifiability calls are made by a competent person; the notebook drafts and structures, and every draft's "For human review:" list is resolved by a human against source records.
• Do not submit feedback from this notebook on a consumer account, and prefer a Workspace account for any work use; confirm your edition's protections with your administrators.
• WHS duties vary by jurisdiction; verify every duty and code reference against your jurisdiction's current instruments before anything operational relies on it.
• Check your organisation's AI use policy before setting this up; a personal consumer account is not an approved channel for work material in most regulated workplaces.

About this configuration

Built against: NotebookLM, which Google renamed Gemini Notebook, in the web app as at August 2026; Google's support pages still live under notebooklm addresses. Platform behaviour, editions and terms change; re-check the sources above before relying on any data-handling claim. TheAICommand is independent: no platform vendor paid for, reviewed or influenced this configuration, and we accept no payment for coverage. See ownership and funding.