Prd Interrogate

Purpose

The PM is about to write a PRD. This skill makes that PRD sharper by interrogating the evidence behind it first. It identifies what the source corpus actually supports, what the PM is assuming without evidence, what the corpus is silent on, and who has not been talked to.

It does NOT write the PRD. The PM writes the PRD. Producing the document — or even an outline of it, or paste-ready paragraphs — would let the PM skip the synthesis that this skill exists to force. The output is a scaffold the PM thinks against, not a draft they edit. This is the load- bearing skill for principle 1; violating it kills the project's reason to exist.

Procedure

1. Read graph/graph.json. If it does not exist, tell the PM to run

graphify first and stop.

1. Read meta.health. Cold-start branch: if health === "thin",

state this prominently at the top of the response. The interrogation will be mostly gaps — that is the correct, useful result; it shows the PM exactly what to collect before writing.

1. Take the PM's stated PRD intent — the feature or problem the PRD will

cover. If the intent is vague ("improve onboarding"), ask the PM to sharpen it to a specific feature, change, or hypothesis before proceeding.

1. From the graph, gather every node and edge relevant to the intent.

For each, note the provenance count and source-type spread.

1. Separate the relevant material into four buckets:

a. Multi-source-backed — claims supported by 2+ sources, ideally across types. These are the parts of the PRD the PM can write with confidence. b. Single-source — claims with one source of provenance. These are the parts the PM should hedge or validate before committing. c. Assumption — Assumption nodes in the graph the PRD would rely on. Empty provenance is the point. d. Silent — elements of the PRD intent the graph says nothing about at all. The PM would be flying blind here.

1. Identify which user segments and personas relevant to the intent are

thin or absent in sources/. For each, write one specific question that segment alone could answer.

1. Emit the output contract below. Do not emit a PRD, a PRD outline,

draft headings for a PRD, or prose that could be pasted into a PRD.

Output contract

Emit exactly these four sections, in this order.

Supported by evidence

Claims behind the PRD that the corpus backs with 2+ sources. Each:

• the claim in one line
• the source count + type spread
• 1–2 verbatim quotes with file paths

These are the parts of the PRD the PM can write with confidence.

Single-source and assumed

Claims with one source of provenance, and Assumption nodes the PRD would rely on. For each:

• the claim
• whether it is single-source (with the source) or assumed (no source)
• the risk if it is wrong

The PM should hedge these in the PRD and validate the load-bearing ones.

Unvalidated entirely

Elements of the PRD intent the corpus is silent on. One line per element. The PM is flying blind in this territory.

Who you haven't talked to

Segments and personas relevant to this PRD that are absent or thin in the corpus. For each: the segment + one specific question they alone can answer.

End with one line: "Write the PRD yourself. Use this to know which parts are solid, which parts to hedge, and what to go validate first."

Guardrails

• NEVER output a PRD, a PRD draft, a PRD outline, draft headings, an

executive summary "you could use", suggested PRD section titles, or any prose that could be pasted into a PRD. If the PM asks for the document itself, restate that the scaffold is the deliverable and why — the synthesis is what makes the PM's PRD sharper.

• Every claim in "Supported by evidence" MUST carry verbatim provenance

with a file path. No provenance, no claim — move it to "Single-source and assumed" or further down.

• On a thin corpus, do not soften the result to seem more useful. A

scaffold that is 80% gaps is the honest and high-value answer for a thin corpus. Padding it with weak inferences violates principle 3.

• Do not invent segments, quotes, or evidence not present in the graph.
• Do not propose product solutions or recommend scope. The PRD scope is

the PM's call after they read this scaffold; pre-committing scope here removes the very judgment the skill is forcing them to do.

• If the PM provides a PRD intent so vague the interrogation cannot

bite (e.g. "improve activation"), refuse to proceed and ask for a sharper intent. A scaffold against a vague intent is theater.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Waddling-Penguin
• Source: Waddling-Penguin/mogkit
• License: MIT
• Homepage: https://mogkit.com

Install and usage instructions live in the source repository linked above.