Thrink — Product Requirements Document

Project managers spend roughly 70% of their week translating information between tools. Slack threads become Jira tickets. Jira becomes Notion docs. Notion becomes status decks. Status decks become emails that nobody reads. The actual unit of value — the decision — takes five minutes. The translation around it takes five days.

Meanwhile, executives owning portfolios of 20–40 projects have one question they can't answer in under a week: am I making money on this? The data exists across eight tools, four currencies, and three time zones. No dashboard reconciles them. The ones that try are read-only museums of stale rollups.

The category response to this has been to build better dashboards. We believe the dashboard is the wrong primitive. The right primitive is a conversation with a system that already knows the project — and answers in plain English.

Goals

• 01Replace the dashboard with a conversation. Default surface is a chat box, not a Kanban.
• 02Two altitudes, one product. Workspace Mode for the team. System Mode for the portfolio exec.
• 03Mobile-first. If it doesn't work one-handed in an Uber, it doesn't work.
• 04Specialist agents over a single super-agent. Nine focused agents share context but specialize.

Explicit non-goals (v1)

• 01We are not building a Jira import tool. Migration friction is intentional — Thrink is the new system, not a wrapper.
• 02We are not shipping a desktop-first experience. Desktop is the export, not the product.
• 03We are not building integrations with every PM tool. Slack, GitHub, calendar — that's it for v1.

Success metrics

• 01Activation: a PM completes their first agent-driven standup within 7 days of signup.
• 02Retention: D30 chat sessions per active workspace ≥ 4/week.
• 03Trust: agent-drafted artifacts (briefs, emails) accepted without edit ≥ 35% by week 4.
• 04Portfolio: System Mode users return to the FX-normalized P&L view ≥ 2× per week.

Thrink is a chat-first project management platform. The user interacts with one chat box and nine specialist agents — Brief, Risk, Schedule, Budget, Stakeholder, QA, Standup, Escalation, Specialist — each addressable by name. Two modes: Workspace (one team) and System (the portfolio). Same agents, different scope.

Decision 01 — Specialists, not a super-agent

Decision: Nine specialist agents with shared context, addressable individually.
Rejected: Single GPT-style super-agent that handles everything.
Why: Specialists are debuggable, evaluable, and replaceable. When Risk gets a bad eval, we replace Risk — not the whole product. Super-agents are opaque and regress unpredictably.

Decision 02 — Two modes, one brain

Decision: Workspace Mode and System Mode share the same agent roster but operate at different altitudes.
Rejected: Two separate products (PM tool + portfolio analytics).
Why: Splitting them means the executive view drifts from the team view within a quarter. Same agents reading the same source of truth means the CFO and the PM are looking at the same number, just zoomed differently.

Decision 03 — Mobile is the product

Decision: Tink, the mobile PWA, is the canonical Thrink experience. Desktop is read-mostly.
Rejected: Desktop-first with a "responsive" mobile site.
Why: PMs do real work between meetings, in Ubers, at school pickup. Standups and approvals happen on phones. Optimizing for desktop optimizes for the wrong moment.

Decision 04 — FX-normalized at the edge

Decision: All cost rollups normalize to a workspace base currency at write-time, with FX timestamped on every transaction.
Rejected: Convert at read-time using a daily mid-rate.
Why: Read-time conversion makes historical P&L drift. CFOs reject numbers that change between two views. Write-time + timestamped FX gives auditable, stable rollups.

• 01Agent eval debt — nine agents means nine eval suites. We're starting with Brief and Risk and rolling out.
• 02Trust calibration — agents that overpromise burn the product faster than agents that underdeliver. Default tone is conservative.
• 03Cold-start in Workspace Mode — new workspaces have no project history to ground agents. Plan: structured onboarding interview that becomes the first source of truth.
• 04Open question — should Specialist agent be domain-tunable per workspace, or is one general specialist enough? Leaning tunable.
• 05Open question — how much do we expose the agent prompts to power users? Transparency vs. brand voice tension.

Rollout

• 01Phase 0 — Design partner cohort (4 PMs, 2 paying). Manual onboarding, weekly feedback loop.
• 02Phase 1 — Closed beta (≈25 workspaces). Workspace Mode only. Tink mobile shipped.
• 03Phase 2 — System Mode opens to portfolio execs in design-partner orgs.
• 04Phase 3 — Public beta with self-serve onboarding and the structured workspace interview.

Eval plan

• 01Per-agent eval set, refreshed monthly with real (anonymized) workspace data.
• 02Acceptance-rate tracking on every drafted artifact (brief, email, summary). Target ≥ 35% accepted-as-is.
• 03Weekly red-team session — design partner sends three intentionally-tricky prompts; we score and triage.
• 04Quarterly trust audit — sample 100 agent outputs, label for accuracy / hallucination / tone, publish internal report.