Aligned Interviewer
Aligned Interviewer — AI-led interviews for qualitative research, built on
Anthropic's Claude models — three stages, end to end:
- Plan — you give a topic and goals; Claude drafts an editable interview guide.
- Interview — Claude conducts a real-time, adaptive, one-question-at-a-time
conversation grounded in the guide, in the respondent's own words.
- Synthesize — Claude analyses the transcript against the guide: emergent
themes, illustrative verbatim quotes, and objective-by-objective coverage.
This is the demo build: a single Next.js app, state held in the browser, no
database — enough to run a convincing live interview and show stakeholders the
whole loop. Scale, persistence, consent flows, and a rigorous analysis pipeline
are deliberately out of scope (see docs/plan for the roadmap).
📄 First live run: see docs/first-test-run.md —
a real end-to-end interview (anonymized) showing the interviewer probing,
following emergent threads, and producing a full synthesis.
Stack
- Next.js (App Router) + TypeScript + React 19
@anthropic-ai/sdk in server-side API routes- Models: Sonnet 4.6 for snappy live interview turns, Opus 4.8 for the
plan and synthesis stages
Models & cost
Models are env-configurable (server-side), defaulting to Sonnet 4.6 for live
interview turns and Opus 4.8 for plan + synthesis:
| Env var | Default | Used by |
|---|
MODEL_INTERVIEW | claude-sonnet-4-6 | live interview turns |
MODEL_REASONING | claude-opus-4-8 | plan + synthesis |
Set MODEL_INTERVIEW=claude-haiku-4-5 for cheap plumbing tests, then switch back
for a real demo. (effort is auto-omitted on Haiku/Sonnet-4.5, which don't
support it.) The interview route uses prompt caching — the interviewer
system prompt and conversation prefix are cached across turns, so cost stays low
even though each turn resends the growing transcript. A full demo interview costs
roughly $0.20–0.40 on the defaults; pennies on Haiku.
Run it
cp .env.local.example .env.local # then add your ANTHROPIC_API_KEY
npm install
npm run dev # http://localhost:3000
Access gate
Set APP_PASSCODE to require a shared passcode for the whole app — enforced
in middleware.ts, so it protects the API routes (the cost surface), not just
the UI. Visitors get a lock screen; the API returns 401 without a valid cookie.
Leave APP_PASSCODE unset to disable the gate (it fails open, so you can't lock
yourself out before configuring it). To enable on the deployed demo, add
APP_PASSCODE in Vercel → Settings → Environment Variables and redeploy.
Deploy (shareable URL)
Configured for Vercel via vercel.json (Next.js framework, lhr1 region —
London, close to UK/Europe respondents; change it in vercel.json if your
audience is elsewhere).
npm i -g vercel # if you don't have the CLI
vercel # first run links/creates the project
vercel env add ANTHROPIC_API_KEY # paste your key (Production + Preview)
vercel --prod # deploy
Or connect the GitHub repo in the Vercel dashboard and add ANTHROPIC_API_KEY
under Project → Settings → Environment Variables.
Plan note: the API routes set maxDuration = 60, which fits the Vercel
Hobby plan. Streaming responses work on Vercel's Node runtime. On Pro you can
raise maxDuration in the route files if long transcripts ever need it.
Other services in this repo
training-log/ is a separate, standalone service (own package.json) that
ingests strength-training session files from Google Drive into a Google
Sheet — unrelated to the interviewer app, just hosted alongside it. See
training-log/README.md.
Layout
app/
page.tsx # stage router: plan → interview → synthesize
layout.tsx, globals.css
components/ # PlanStage, InterviewStage, SynthesisStage, chat UI
api/
plan/route.ts # Opus → structured interview guide
interview/route.ts # Sonnet → streamed interviewer turns (SSE)
synthesize/route.ts # Opus → structured analysis
lib/
anthropic.ts # SDK client + model constants
prompts.ts # interviewer persona + plan/synthesis prompts
types.ts # shared Guide / Transcript / Synthesis types
