How do I show AI skill on a PM resume without sounding fake?

Name the tool, the task, the verification step, and a business number. Example: built a PRD review SOP that cut spec-review cycle time from 4.5 days to 2.1 and caught 11 missing-metric defects over 8 launches.

Claude or Gemini for PRD review?

Use both on the same packet. A finding counts if both models surface it or a human confirms it. Do not treat a single polished draft as a review.

Workplace SOP · not a prompt dump

Product Manager AI Transition Guide

A four-step PRD review SOP with an input schema, guardrails, a 16-point rubric, and a 30-day matrix from L2 operator to L4 eval owner. Built for senior ICs who need proof a manager can check.

30-day matrix Get your starting level

Direct answer

An AI-enabled product manager does not “write PRDs with ChatGPT.” They bind a packet (problem, evidence, metrics, constraints), run a review agent that is forbidden to invent users or dates, gate P0s with engineering or design, and log cycle time plus defect catch. If you cannot name the verification step and a number, you do not have a career artifact yet.

How we test this SOP. Same packet through Claude and Gemini. A finding counts only if both models surface it or a human reviewer confirms it. The finding log below is a worked example, not a named case study. We will replace it when a practitioner signs real hours. See the editorial policy.

AI-enabled product manager, defined

A PM who can choose a model, structure messy input, score output against a rubric, and refuse to ship invented scope. They are not an ML engineer. They are the person who turns discovery noise into reviewable artifacts: PRD outline, stories, edge cases, events, launch risks.

Competency scale (same L1–L5 as the survey)

Level	Name	What they can ship	What they must not claim
L1	Prompt user	One-off drafts with heavy rewrite	Workflow ownership
L2	Workflow operator	Saved prompts + checklist on a repeating task	Agent / automation skill
L3	SOP builder	Schema + rubric + human gate on one job	Org-wide transformation
L4	Eval owner	Quality bar, failure taxonomy, before/after metrics	Model development
L5	Multi-agent orchestrator	Routed agents, memory, connectors, governance	Research-scientist role

PRD review agent in 4 steps

Bind the packet. PRD draft, tickets, constraints, success metrics. Output is validated JSON. Done when the schema validates and there are no unnamed “stakeholders.”
Run the review agent. System prompt + packet + rubric into Claude or Gemini. Output is a finding log with severity and evidence spans. Done when every finding cites a packet field, not vibes.
Human gate. PM + eng/design lead. Accept / rewrite / reject. Done when no P0 remains open without an owner.
Ship and measure. Decisions + before/after PRD. Output is a review memo and cycle-time log. Done when hours saved and P0 catch rate are recorded.

Input schema

Reject the run if any required field is empty.

prd-review.schema.json

{
  "prd_id": "string",
  "problem": "string",
  "user": "segment + job-to-be-done",
  "success_metrics": ["leading", "lagging"],
  "non_goals": ["string"],
  "constraints": {
    "eng_weeks": "number",
    "platforms": ["ios", "web"],
    "compliance": ["PII", "SOC2"]
  },
  "open_questions": ["string"],
  "evidence": ["interview notes", "ticket ids"]
}

System prompt rules

prd-review.system.txt

You are a staff product reviewer. You do not invent users, metrics, or engineering effort. If evidence is missing, emit a P0 finding with gap: evidence instead of a polished requirement.

Never convert an assumption into a committed scope item. Flag roadmap language that implies a date or capacity the packet does not contain. Quote the source span for every finding.

Score the packet with the 8-row rubric (0–2). Ship threshold: no P0, total ≥ 12 / 16.

Review rubric (0–2 per row)

Ship threshold: no P0, total ≥ 12 / 16.

Check	0	1	2	If fail
Problem is evidenced	Opinion only	One source	≥2 independent sources	P0
Success metric is measurable	Vanity / missing	Lagging only	Leading + lagging + window	P0
Non-goals are explicit	None	Vague	Named cuts with owners	P1
Constraints bound the solution	Ignored	Listed	Tied to scope cuts	P1
Edge cases and empty states	Happy path only	A few	Errors, permissions, offline	P1
Analytics events named	None	Page views	Funnel events + properties	P1
Open questions have owners	Hidden	Listed	Owner + due date	P2
Assumptions labeled	Written as facts	Mixed	Tagged + test plan	P0

Worked example — finding log

Synthetic packet for teaching. Not a signed case study.

ID	Finding	Evidence span	Rewrite	Decision
P0-1	Success metric is “increase engagement”	Goals §2	Raise 7-day retained teams from 22% to 30% in 90 days	Accept
P0-2	Enterprise SSO listed as in-scope without eng estimate	Requirements R12	Move SSO to non-goal; spike in Q4	Accept
P1-3	No empty-state copy for zero projects	UX § missing	Add empty, loading, and permission-denied states	Rewrite
P2-4	Open question on pricing has no owner	Risks §4	Assign to Finance PM; decision by 12 Sep	Accept

Claude vs Gemini for spec review

	Claude	Gemini	Rule
Long packets	Strong at citing spans	Strong at structured tables	Dual-pass; keep intersection + human confirms
Invented metrics	Still happens if schema is empty	Still happens if schema is empty	Fail closed on empty success_metrics
When to skip both	Live strategy calls, unannounced headcount, customer legal language you have not read		Human only

30-day matrix for senior PMs

This is what the homepage quiz should return. Pick the level that matches your score (under 30% → L2, 30–59% → L3, 60%+ → L4).

Portfolio template

Resume bullet — proof vs puff

Before

Used ChatGPT to help write PRDs and improve documentation quality across the product team.

After

Built a PRD review SOP (schema + rubric + human gate) that cut spec-review cycle time from 4.5 days to 2.1 days and caught 11 missing-metric defects over 8 launches.

Rule: tool + task + verification + business number. No number, no bullet. Use redacted real work only.

FAQ

What is an AI-enabled product manager?

Someone who binds input, runs a gated review, and can explain what the model is not allowed to invent. Not a prompt collector.

Do I need to code?

Not for L2–L3. L4 needs an eval sheet (even a spreadsheet). L5 needs connectors your security team has approved.

Where are paste-ready product prompts?

In the prompt library. Use them inside this SOP, not instead of it. Persona landing: Product.

Start with your level, then one packet

Quiz for the 30-day plan. Survey if you want the industry benchmark to include people like you.

Take the 8-question quiz Wave 1 survey