Inline advice pills — colored badges inside AI answers

When Inline Advice is toggled ON (topbar Settings → Inline advice), the AI annotates key phrases inside its answer with colored pills tying each phrase back to a source. Six pill types, each a different color + icon. Click off the toggle and the AI emits clean prose with no pills.

The 6 pill types

Pill	Color	Icon	What it means
🟣 resume	violet	📄 file-text	This phrase came from YOUR resume. Highlights a specific role, company, project, or skill the AI is pulling from your uploaded PDF.
🟦 memory	indigo	🧠 brain	This phrase pulls from your persistent memory — your name, school, location, or other facts the AI remembers across sessions.
🟢 jd	emerald	🎯 target	This phrase ties to the job description you pasted. The AI is matching your background to a specific requirement.
🩷 drill	rose	🛠️ drill	A practice prompt — the AI suggests you drill this answer further because it's a likely follow-up.
🩵 askback	cyan	❓ question	A question to ask the interviewer back at the right moment, woven into the answer.
🟡 coach	amber	💡 lightbulb	A coaching tip — meta-advice on how to deliver this answer (e.g., "pause for emphasis here").

Why pills exist

Without pills, the AI answer is a black box — you can't tell what it's pulling from. Pills let you verify:

• Is the AI hallucinating that I worked at Google? → No resume pill = no source, treat with suspicion.
• Is the AI tailoring to this JD? → Look for jd pills. If none, paste the JD into Job Context.
• Did the AI remember my CMU master's? → memory pill on "CMU" = yes, it's grounded.

The pills are your trust signal that the answer is grounded in your real data, not invented.

How it works under the hood

The AI is given an <annotation_rules> block in its system prompt that says: "when you reference something from the user's resume, wrap it like {{resume:phrase}}. When you reference memory, use {{memory:phrase}}." The frontend then parses those tags out of the markdown and renders them as the colored pills you see.

If a pill renders incorrectly (wrong type, weird phrase), it usually means the AI emitted a tag with a typo — {{ resume :foo}} (with extra spaces) won't match. We've seen this on smaller models (Haiku, GPT-4o-mini) about 1% of the time. Sonnet/Opus get it right ~99.5% of the time.

When pills are most useful

• Behavioral round. The interviewer asks about a specific role. Pills tell you which company-bullet the AI is pulling.
• Technical screen with a panel. Watch for coach pills mid-answer suggesting "slow down here" — they're calibrated for technical-screen pacing.
• JD-tailored prep. Run a session WITH the JD pasted in Job Context. Look for jd pills to see exactly which requirements the AI is matching to your background.
• Self-audit your resume. Run 5 questions, watch the resume pills. If certain bullets never get cited, they're probably weak — rewrite them.

When pills can distract

• High-stakes live interview. First time using inline advice in a real round = cognitive overload. Practice with it ON in mock rounds first.
• Coding round. The AI mostly returns code blocks. Pills inside code look weird. Toggle OFF for coding.
• You're reading the answer aloud verbatim. Don't. Pills are for grounding signals, not for screen-reader pacing.

Pills + Coaching mode

Both can be on simultaneously — they cover different moments:

• Coaching mode = nudges DURING YOUR speaking turn (live transcript pattern detection)
• Inline advice pills = annotations INSIDE the AI's answers (LLM-generated structured tags)

They don't conflict. Many users run both ON during practice + behavioral interviews, both OFF during coding rounds.

Gotchas

• Pills cost ~2-5% more tokens per answer. The AI emits extra characters for the tag syntax. Counts against your daily AI quota on Free / Hour Pack.
• Pills disappear in copy-to-clipboard. When you click Copy Answer, the clipboard gets the plain text without pill markup. So if you paste the AI answer into a notepad to study, the source citations are stripped.
• No exports. Transcripts (Pro+ feature) include pill text but not pill colors/icons. Plain text only.

Where it lives in code

apps/web/app/interview/_components/inline-advice.tsx — the AnnotatedMarkdown renderer. AI system prompt at apps/api/domains/ai_orchestrator/prompts/system_prompts.py — <annotation_rules> block injected when inlineAdvice=true.

Still stuck?

Help & Feedback → Contact us with category "Something is broken" + screenshot showing the pill that's wrong + what you expected.

Video walkthrough

Coming soon.