Coaching mode — AI nudges during your turn

Topbar Settings (gear icon) → toggle Coaching tips. When on, a rule-engine fires small nudges in the AI Coach panel during your speaking turn (not the interviewer's). Off by default — opt-in feature.

What it does

While you're answering a question, the AI watches the live transcript for patterns that often hurt candidates and surfaces a tiny nudge in real time. Examples:

• You've been talking for 90+ seconds → "You're at 90 seconds. Wrap with a concrete result line."
• You haven't mentioned a metric → "Ground this with a number — how many users, what % improvement."
• You said "we" 5+ times → "Switch from 'we' to 'I' — interviewer wants YOUR contribution."
• You filler-word-stacked ("um", "like", "you know") → "Three filler words in 20 seconds. Pause and reset."
• You jumped to action before context → "Add Situation before Action — they need the setup."

Each nudge is one line, glanceable in <1 second. Nothing scrolls past — just the latest one stays visible until the next fires or you stop talking.

When to leave it ON

• Practice sessions — learn your verbal tics without an interviewer's pressure
• Recorded mock interviews — watch the playback later with nudges as breadcrumbs
• Long-form behavioral rounds — easy to ramble; nudges keep you tight
• High-stakes screens where you tend to rush — the "slow down" nudges are especially good

When to turn it OFF

• Live job interview, first time using it — context-switching to read a nudge while talking is a learned skill. Practice with it first.
• Coding round — coaching is calibrated for verbal patterns; in a coding round you're typing more than talking. Less useful.
• Anxiety-sensitive moments — for some people, the nudges feel like another voice judging them. Off until you're comfortable.

Where the nudges come from

Rules in apps/api/domains/coaching/ evaluated against the rolling transcript window. NOT a full LLM call per nudge — these are deterministic patterns ("if X then Y") tuned from interview-prep coaches. Why deterministic instead of AI-generated:

• Fast enough to fire in real time (sub-100 ms vs ~300 ms for an LLM round-trip)
• Predictable — same input always produces same nudge, makes it teachable
• Free of inference cost — runs even on free tier

Inline advice pills are different

There's a separate setting: Inline Advice (also in topbar Settings). That feature surfaces colored pills inside AI answers (e.g., 🟣 resume, 🟢 jd) tying suggestions back to specific data sources. See Inline advice pills for that. Both can be on simultaneously — they cover different moments (your turn vs the AI's answer).

Gotchas

• Nudges fire on transcription, not your meaning. If the AI mishears you, the nudge can be wrong. Especially on accented speech with poor mic — verify Speechmatics is hearing you correctly via the live transcript panel.
• No nudges = silence is OK. Some answers don't trigger any rule. That's the rule-engine choosing not to interrupt — not a bug.
• Toggling mid-session is fine. No state carryover; the rule engine just stops watching.

Still stuck?

Help & Feedback → Contact us with category "How do I…?" — also tell us what behavioral nudge you wish existed but doesn't. The rule list grows from user feedback.

Video walkthrough

Coming soon.