Anchor

Practice the conversation before it costs you.

A negotiation simulator that puts you across the table from an AI opponent that actually pushes back — one that anchors low, stalls, manufactures deadlines, and will walk away from you. When it's over, it hands you a report that quotes your own messages back to you and tells you exactly where you gave ground you didn't have to.

Built with Next.js 16, Supabase and Gemini, running entirely on free tiers.

Open the live demo — one click, no signup, into an account that already has negotiation history in it.

---

The problem

Negotiation is a skill you almost only practise live, in the situations where getting it wrong is most expensive. There is no low-stakes rehearsal for asking for a raise. Worse, the feedback loop is broken at both ends: you rarely find out what the other side would have accepted, and generic advice ("be confident", "anchor high") is impossible to act on in the moment.

Two things had to be true for this to be worth building:

1. The opponent must be genuinely hard to beat. An LLM asked to role-play a tough negotiator is agreeable by default. It concedes because conceding is what helpful assistants do. If the AI folds, the practice is worthless.
2. The feedback must be specific to what you actually said. "You could have anchored higher" teaches nothing. "At message 4 you said I could probably come down to $165,000 — you gave away $10,000 before they had asked" teaches something.

Everything below follows from those two constraints.

---

How it works

The opponent cannot decide to concede

Each turn, the model returns structured JSON: its reply, the tactic it read in your message, and the position it wants to move to. Application code then clamps that position into a range the persona's playbook allows.

model proposes  →  ledger clamps  →  user sees the clamped number

Conceding past the reservation value is arithmetically impossible rather than merely discouraged. No amount of persuasion — or instruction-shaped text pasted into a message — moves an opponent further than its budget permits.

Two things are deliberately withheld from the model:

• its own reservation value. A model that knows its floor will eventually allude to it, drift toward it, or hand it over when asked nicely. It is told only how far it may move this turn.
• your walk-away number, which you enter privately during setup.

Four opponents that differ mechanically, not tonally

Tone alone produces four models that cave at the same rate in different words. The real differentiation lives in playbooks.ts as numbers the model never sees: how far each will ever move, what makes it move, what it does when you concede, and what makes it leave.

Measured across a 4-turn scripted replay:

user behaviour	Lowballer	Staller	Professional	Closer
holds firm, no justification	0%	0%	0%	walks
concedes unprompted each turn	0%	0%	0%	0%
brings verifiable evidence	15%	7%	8%	25%
turns hostile	walks	stalls on	walks	walks

The teaching signal falls straight out of that table: evidence is the only thing that moves anyone. Repeating your number moves nobody. Conceding unprompted is never reciprocated.

The Friendly-but-Firm Professional is the difficulty-4 opponent, not the aggressive one — it has the smallest concession budget of the four. Losing to someone pleasant feels like winning, which is exactly what makes it hard.

The report has to cite its sources

The report generator is given the transcript with explicit [#n] markers, and its output schema requires a message_seq and a verbatim quote on every observation. Then every citation is verified against the transcript:

• quote found in the cited message → kept
• quote found in a different message → re-pointed at the real one
• quote found nowhere → discarded

If too few survive, the whole report is regenerated with the failures named. A model cannot bluff a citation past this, which is what makes the report worth reading rather than merely well-written.

---

Stack

Framework	Next.js 16 (App Router), React 19, TypeScript
Styling	Tailwind CSS v4
Database	Supabase Postgres with RLS
Auth	Supabase Auth — email/password + Google OAuth
AI	Gemini (3.5-flash-lite for turns, 3.8-flash for reports)
Charts	Recharts
Motion	Motion (Framer Motion)
Voice	Web Speech API — no external service
Hosting	Vercel

Type: Fraunces (display), Instrument Sans (body), IBM Plex Mono — every figure renders in tabular mono so offers and deltas align down the page. Colour carries meaning rather than decoration: brass is you, steel is the opponent who has not moved.

---

Key technical decisions

Multi-tenancy is enforced by Postgres, not by application code. Every table has RLS enabled and every policy is scoped to auth.uid(). Server routes are built on the publishable key plus the user's own cookie, not the secret key — so server-side reads pass through the same policies the browser does. The secret key never touches a user's data: its only runtime uses are the keepalive cron and housekeeping for the shared demo account (below).

Transcripts and reports are append-only at the database level. No UPDATE/DELETE policy exists on either. A transcript you can edit after the fact makes the feedback report meaningless, so the database refuses rather than trusting the app. Once a session finishes it is frozen too: its owner can no longer update it, delete it, or append to its transcript.

This is the beginning of the README. Read the rest at the original. Read the full README ↗