dsh-ai-pm

When using AI to draft product requirement documents, the biggest risk is not speed but fabrication: inferences disguised as facts, the absence of an explicit human sign-off, and silent downstream drift when upstream specs change. dsh-ai-pm plugs the pm-scaffold controlled PRD workflow into DeepSeek Harness, turning "AI writes requirements" into a verifiable loop of AI drafting, self-audit, and genuine human confirmation so that no artifact can be silently overwritten by a machine.

Core capabilities include six-state knowledge annotation (Fact, Decision, Assumption, AI Inference, Unknown, Conflict) that guarantees AI inferences can never masquerade as facts; an un-bypassable human gate ensuring the confirmed state is produced only by an authorized reviewer in the GUI; bidirectional traceability linking objectives, stories, features, and acceptance criteria in both directions; and a change-reflow mechanism that automatically invalidates downstream artifacts when upstream content shifts. On the agent side it exposes model tools such as init, status, entry, gate, reflow, skill, and artifact; on the GUI side (Settings, Plugins, AI-PM) it provides a requirement kanban board and a confirmation interface.

This plugin is well-suited for product teams that rely on AI to draft PRDs but require a product manager or designated reviewer to give final approval, as well as for agent engineers who want to bring structured requirement management into their DeepSeek Harness workflows.