Trainer mode

Trainer mode is for the engineer recording a procedure — the senior whose know-how you're capturing. The Core App records their POV video, transcribes their narration, and after the session the platform extracts a structured procedure from the footage.

When to use trainer mode

• Capturing how your most experienced engineer does a job — a boiler service, a fault diagnosis, a customer-facing escalation — so the procedure becomes reusable training content.
• Creating evidence of a one-off "good outcome" that can be reviewed afterwards (compliance audit trail, dispute resolution, training material).
• Building your tenant's procedure library — every trainer session adds to the corpus your trainees draw on.

Each trainer session becomes one or more procedures after the post-session processing pipeline runs.

What the engineer experiences

1. Put the glasses on, power them on. The Core App authenticates automatically.
2. Choose the task (if you have an open task to record against) or start a freestanding session.
3. Say "Start training session."
4. A short pre-flight runs silently (~2 seconds): voice channel up, video stream up, AI participant joined, all checks pass.
5. The session is live. The engineer narrates what they're doing as they do it — naturally, like explaining to an apprentice on the job. The AI assistant listens but stays mostly silent in trainer mode (it isn't actively coaching).
6. When the engineer's done, they say "Stop training session."
7. The session is sealed; the recording uploads in the background.

What's running under the hood

• Voice + audio flows through OpenAI's Realtime API (the gpt-4o-realtime-preview-mini model in trainer mode — lower cost, listen-and-transcribe-focused).
• Video streams continuously from the glasses camera over WebRTC to TrainAR's storage.
• No skills, no tool calls in trainer mode. The trainer agent doesn't search knowledge or run skills — it captures, it doesn't intervene.

After the session

The Core App's post-session pipeline runs on the tenant side. It segments the recording into discrete procedures, generates a name + description for each, scores them, and auto-approves them. By the time the session appears in Dashboard → Sessions, the procedures are already named, clipped, scored, and ready.

Typical processing time: 5–15 minutes for a 20-minute trainer session. See Procedure extraction for what the pipeline produces and where it shows up.

Tips for good trainer sessions

Narrate what you're doing, not what you're thinking. "Now I'm taking the front cover off and checking the pressure gauge" is more useful than "this might be a pressure issue." The extracted procedure works better when narration tracks visible actions.

Pause when you move between tasks. A clear two-second pause between distinct steps helps the procedure-extractor segment cleanly. You don't need to say "next step" — just pause.

Show, don't just describe. If you're checking a fault code, look at the screen for two beats so the camera captures it cleanly. The post-session vision pass picks up text from the footage; clean shots improve accuracy.

Talk about safety as you go. "I'm isolating the gas at the meter before I do anything else." Spoken safety callouts surface in the extracted procedure's description, so trainees following the same procedure later get the same context.

Keep sessions under 30 minutes when you can. Long sessions process slower and produce longer procedures that are harder for trainees to follow as a single unit. Two 15-minute sessions usually beat one 30-minute one.

What you don't do

You don't need to:

• Press a button to capture video — it's continuous from session start.
• Tell the AI what skill to use — it doesn't run skills in trainer mode.
• Annotate or label anything during the session — extraction is automatic.
• Approve the procedure afterwards — auto-approval happens on the tenant side.

Viewing the session afterwards

The session and its extracted procedures appear in Dashboard → Sessions automatically. The detail page is read-only — see the recording, the auto-generated summary, and the procedures the pipeline pulled out (each as a named sub-clip with a thumbnail). There's no per-step editor, no approve/reject step, no publish toggle in the Dashboard.

See Viewing a session and Procedure extraction for the full view side.