Briefer

A Hebrew-speaking Socratic voice mentor for first-year Visual Communication students at Bezalel. Upload the assignment brief, and it walks you through your own thinking — out loud, on the move, without ever handing you the answer.

Design students get stuck between receiving a brief and having a first idea. A chatbot that answers for them removes exactly the part of the work that teaches. Briefer is built the other way round: it is a voice mentor with a hard rule against giving solutions, so the student keeps authorship of every idea while still having somebody to think against — walking to campus, standing in a shop, wherever the thinking actually happens.

Built as a research-through-design project for the Interaction Design studio, grounded in interviews with first-year Visual Communication students.

The roadmap. The uploaded brief is parsed into ordered stations — research, ideation, planning, execution, review — that the mentor follows linearly, one at a time. The student owns the plan: steps can be added, reopened, or skipped, and the mentor names any unfinished earlier step rather than quietly moving on. The voice stage. A Gemini Live session with idle / listening / speaking states, transcribed both ways so every spoken turn persists as chat history. Text chat is the fallback in places you can't talk.

The Socratic constraint

server/mentorPrompt.ts is the single source of truth for mentor behaviour — the same rules feed brief analysis, text chat and the live voice session, so the mentor can't drift between modes. The rules are absolute, not suggestions:

• No solutions, no options, no binary choices. Not even a helpful-looking "is it sharp and bouncy, or flowing?" — a question like that injects the answer's vocabulary. The mentor asks the student to supply the description instead.
• One open question per turn. A clarifier on the same point is allowed; a second question is not.
• No proactive interpretation. The mentor never unpacks a concept, metaphor or image on the student's behalf — that is the assignment.
• Explicit consent before touching state. Notes and roadmap steps change only after the student says so out loud.
• Nothing invented. No facts, names or requirements that aren't in the brief.

How it works

brief PDF ──▶ Gemini analysis ──▶ roadmap of stations
                                        │
student speaks ──▶ Gemini Live (voice, both-way transcription) ──▶ spoken reply
                          │
                          └── tool calls: add_note · update_note · delete_note
                                          complete_step · add_step · reopen_step
                                  ↓
                    relayed over WebSocket → applied to project state → confirmed aloud

Piece	What it does
server.ts	Express + ws. Holds the Gemini Live session, relays tool calls to the client and results back, serves the Vite app.
server/mentorPrompt.ts	The Socratic rules, brief-analysis prompt, chat and voice system instructions — one module, three consumers.
src/hooks/useVoiceAgent.ts	Client side of the voice session: audio in/out, transcription, executing tool calls against project state.
src/hooks/useProjects.ts	Projects, roadmap, notes and the cross-conversation learning profile, persisted in localStorage.
src/components/	Voice stage, progress bar, roadmap view, notes and conversation drawers, upload modal.

Pausing rebuilds the whole context on resume — brief, roadmap state, notes, conversation history, learning profile — so a conversation picked up hours later doesn't start from nothing.

Design

Hebrew, RTL throughout, mobile-first with the drawers docking as permanent panels above 1024px. The visual language is modern Bauhaus: red, blue, yellow, black and off-white, a strict grid, flat geometric shapes with hard offset shadows, Rubik, and micro-animations via motion.

Running it

npm install
cp .env.example .env.local     # add GEMINI_API_KEY
npm run dev                    # tsx server.ts → http://localhost:3000

Google Cloud Text-to-Speech needs GOOGLE_APPLICATION_CREDENTIALS pointing at a service-account key; TUNED_CHAT_MODEL, VERTEX_PROJECT and VERTEX_LOCATION are optional and switch text chat to a fine-tuned model on Vertex AI.

npm run build bundles the client with Vite and the server with esbuild; npm start serves the result.

Status

A studio project, not a product: state lives in localStorage, there are no accounts, and it's been used by its author and a handful of students rather than a cohort. The research it came from — interviews, the usability study, the fine-tuning dataset — stays out of this repository.

License

Apache-2.0