Product
One product brain behind every artifact. Below is what runs today, and what’s on the way — marked honestly.
Bring transcripts from Zoom, Teams, Otter, or Fireflies — turned into structured decisions, owners, actions, risks, and PRD-ready snippets, each timestamped to the moment it was said.
Draft PRDs from transcripts or signal clusters with inline citations, then /red-team-prd and /ship-check to stress-test them before they ship.
RICE and ICE today — every component linked to evidence and a plain-English rationale, not a number pulled from the air. More frameworks (WSJF, MoSCoW) on the way.
Human-in-the-loop by construction. Agents propose; a person approves; every decision — and every override — recorded in an immutable audit.
pm.gold is an MCP server. Add it to Claude, ChatGPT or Codex and your agent gets four read tools over the record — find precedent, read an approved decision, fact-check a document, compare intent to what shipped — plus three that can only propose, one of which is research capture: your assistant searches with its own tools and hands us a URL, we fetch the page ourselves, and it enters as evidence awaiting review. A key inherits exactly the permissions of the person who issued it, and there is no approve tool for an agent to call.
A committed decision carries its assumptions, each naming what would prove it false. The pages those point at are re-fetched on a schedule and diffed, and when a cited source moves, the decisions still standing on the old version are found by query rather than by asking a model whether something still holds. Each arrives as a proposal in the review queue with the diff attached — never as a silent edit — and a weekly interrupt budget caps how often it may reach you.
What share of your committed decisions still rest on current evidence — and beside it the negative space: the objectives you hold no evidence about at all. A record can be wrong by being silent, and a health number that only counts what exists would never show it.
Once a week, what is waiting for you and what has gone stale — to your own inbox or your own Slack, on a human click. Counts, titles and links only; never the evidence text, because a channel is a wider audience than the people who could read the source. Email goes only to your own verified address; the Slack destination is set once by an admin and stored, never passed in with a request.
A hypothesis, what it predicted, and what actually happened, recorded against the decision that motivated it. This is what turns precedent recall from "we discussed this before" into "we tried this before, and here is how it went".
Read cores and fail-closed permission mapping are built for SharePoint / M365, Confluence, Jira, GitHub, Slack, Drive and more — each item keeping the access rules it had at the source. Scheduled provider sync is still staged: until it is verified against a real tenant it returns an honest error rather than pretending. Files, email, decks, spreadsheets, archives and web pages all import today.
Supported claims link back to the evidence behind them. Sections with no supporting evidence are flagged as unsupported rather than filled with plausible text — you can see the difference at a glance, and so can your reviewer. That is a stronger guarantee than "everything is cited", because it tells you where the record is thin.
Committed decisions and their outcomes come back when your team decides something similar, so a mistake already made is visible before it is repeated. This is designed to compound with use; we instrument it and publish what we measure rather than a marketing number.