A challenge I set myself · Product ops · 2026

Raw customer feedback in. A prioritized, spec'd roadmap out.

4 sourcesGong, Zendesk, App Store, NPS → one pipeline
~$17M/yrTAM the top theme unlocks (illustrative)
1 eveningspec'd as a PM, built with AI
The business case

Product teams drown in feedback, and prioritization ends up living in a few overloaded heads. A designer once put the real gap to me directly: there's no voice of reason on scope — everyone says "great idea, let's build it," and nothing gets cut. So I built the voice of reason. This turns scattered feedback into a ranked, revenue- and TAM-sized, spec'd roadmap — with a human approving every step.

The framing that matters: it doesn't just say "this fixes a pain." It says "build this and ~$17M/yr of a $68M segment becomes winnable." Fixes become markets. Illustrative data, real architecture.

What I built

Three parts, one brain: a governed MCP server (the deterministic engine), a Claude skill (the workflow), and an interactive triage tool (below — go click it). One rule keeps it honest, and it's the part an engineering leader cares about most:

AI owns the judgment. A deterministic engine owns every number. And the one action that writes is gated and eval-tested. RICE and revenue can't drift between runs, the model has to show its work, and a spec never gets filed without an explicit confirm — a guardrail with its own test so it can't quietly regress.

Try it — it's live

Pick a piece of feedback, see whether other customers echoed it, set your judgment on the sliders, and watch it reason to a verdict. It deliberately won't green-light a single loud voice.

Trouble in the frame? Open it full-screen ↗

How it works

1 · Where it comes from. Real sources — Gong calls, Zendesk tickets, App Store reviews, NPS — normalized into one shape. Fixtures today; going live is a config change, not a rewrite.

2 · Triage one signal. A customer says something; it resolves the account, checks how many others said it (pattern vs. one voice), scores it, and gives a verdict: prioritize now, gather more evidence, or backlog.

3 · Rank everything. It clusters all the feedback into themes and ranks them by RICE and revenue-at-risk — with the reasoning for each, and a plain-English read of what the RICE number even means.

4 · Draft the spec. The top theme becomes a decision-ready PRD — and I approve the write, it doesn't.

The output

The PRD it files isn't a wall of text. Every claim links back to the actual call or ticket it came from. The RICE score comes with a plain-language reading. There's a TAM section sizing the market the work unlocks, the four product risks (value, viability, usability, feasibility), a rollout that de-risks from shadow to automate, and open questions. Fourteen sections, all skimmable.

"The best prioritization isn't an opinion — it's a number you can trace back to a customer." AI owns the judgment; a deterministic engine owns every number.