One structured intake, one synthesis, one verify pass. User stories, interface copy, scope table, decision log and acceptance cases — in your template, in minutes. Asset-risk side effects are written as hard rejections.
Hedy writes PRDs as a governed, on-premise AI employee, not a cloud assistant. It runs a five-question intake, synthesizes a structured §0-6 draft, switches to engineering mode to fill technical contracts, then to review mode to challenge gaps. Any action that risks real assets — deleting docs, overwriting tickets, mass-messaging — hits a default-deny whitelist and hard-refuses. One Hedy owns the whole PRD workflow, self-hosted, with append-only audit and cited sources.
Last updated: 20 July 2026
Problem, entry point, minimal scope, touched systems, references. One round, most of the signal.
From background through acceptance cases — with scope stated as "we will not build" pairs, each with a reason.
On request: auth chains, error codes, idempotency and transaction boundaries — checked against a hard list.
Point Hedy at an existing PRD and get findings ranked by severity, referenced by section.
Most AI writing tools stop at drafting prose. Hedy runs the full PRD workflow as one employee. It opens with a five-question intake — problem, user, scope, constraints, success metric — instead of guessing. Then it synthesizes a structured document across §0-6: context, goals, users, scope, functional requirements, non-goals, and open questions. The output is a real PRD skeleton your team can review, not a wall of text. One Hedy covers what an intake PM, a spec writer, and a reviewer used to split.
The work runs in modes. In synthesis mode Hedy assembles §0-6 from the intake answers. In engineering mode it fills the technical contract — data shapes, API surfaces, edge cases, dependencies, rollout constraints — so the spec is buildable, not aspirational. In review mode it turns adversarial: it re-reads its own draft, flags missing non-goals, ambiguous acceptance criteria, and unstated assumptions. Because Hedy is self-hosted on your infrastructure with your own model keys (including local vLLM or Ollama), the whole PRD never leaves your network and there are no sub-processors.
Governance is the product, not a setting. When PRD work touches real assets — deleting a Confluence page, overwriting a Jira epic, mass-messaging a channel — Hedy does not proceed. Write operations run default-deny against a whitelist, so any unlisted side-effect is a hard refusal, not a soft warning it can be talked out of. This limit lives in config and code, not in a prompt, so it cannot be prompt-injected away. Retrieval respects the asker's ACL, so a PRD assembled for one requester never surfaces documents they can't see.
Every step is accountable. Actions land in an append-only audit log — no edit, no delete endpoint. Higher-risk operations pass through L0/L1/L2 approval tiers, and Hedy's evidence gate requires that factual claims in the PRD are cited verbatim to their source, so requirements trace back to the interview answer or document they came from. Feishu/Lark is a first-class surface alongside Slack, and the deployment is a ~30-minute Docker Compose install licensed per seat with an offline Ed25519 license.