Product Manager Prompts for Generative AI

╔═════════════════════════════════════════════════════════════════════════╗
║                                                                         ║
║   ██████╗ ███╗   ███╗    ██████╗ ██████╗  ██████╗ ███╗   ███╗██████╗ ████████╗███████╗
║   ██╔══██╗████╗ ████║    ██╔══██╗██╔══██╗██╔═══██╗████╗ ████║██╔══██╗╚══██╔══╝██╔════╝
║   ██████╔╝██╔████╔██║    ██████╔╝██████╔╝██║   ██║██╔████╔██║██████╔╝   ██║   ███████╗
║   ██╔═══╝ ██║╚██╔╝██║    ██╔═══╝ ██╔══██╗██║   ██║██║╚██╔╝██║██╔═══╝    ██║   ╚════██║
║   ██║     ██║ ╚═╝ ██║    ██║     ██║  ██║╚██████╔╝██║ ╚═╝ ██║██║        ██║   ███████║
║   ╚═╝     ╚═╝     ╚═╝    ╚═╝     ╚═╝  ╚═╝ ╚═════╝ ╚═╝     ╚═╝╚═╝        ╚═╝   ╚══════╝
║                                                                         ║
║   119 practical prompt assets for AI-assisted product management        ║
║   ChatGPT • Claude • Copilot • Gemini • and others                      ║
║                                                                         ║
║   Community Build v2.5 • August 9, 2026 • CC BY-NC-SA 4.0               ║
╚═════════════════════════════════════════════════════════════════════════╝

Practical AI tools for both the strategic thinking and the daily execution of product management — that also teach you to prompt better while you use them.

Every asset here works by copy-paste into ChatGPT, Claude, Gemini, Copilot, or any AI assistant. No installs, no accounts, no framework. Pick the problem you're solving, grab the prompt, go.

(Looking to equip an AI agent rather than a chat session? See the companion Product Manager Skills repo.)

---

Start With Your Problem

"I need this artifact, and I have my context ready"

You know the situation; you want execution-quality output in one pass. Go to /prompts/:

• Write a PRD from your discovery notes → prd-prompt-template
• Understand what customers actually need → jobs-to-be-done
• Turn win/loss debriefs into ground truth → win-loss-analysis
• Map who matters and how to engage them → stakeholder-map
• Decode what an inbound ask really wants → incoming-request-breakdown
• Kill the launch on paper before reality does → premortem
• Lose the legacy product without losing the legacy customer → EOL suite
• Write user stories engineers don't hate → user-story
• Frame an initiative as testable hypotheses → lean-ux-canvas
• Design an agentic AI system responsibly → agent-strategy-canvas
• Save this session's context for the next one → session-saver

"I need to think it through — facilitate me"

You have a fuzzy situation and want a guided working session that ends in the finished deliverable. Go to /workshops/:

• Competitive battle card → battle-card-workshop
• PRD, section by section with checkpoints → prd-workshop
• Opportunity solution tree + first experiment → opportunity-solution-tree-workshop
• Build / don't-build business case with the math shown → feature-investment-workshop
• Problem statement + "How might we" → problem-framing-canvas-workshop
• Structured user-pain research table → painstorming-workshop
• Product/feature sunset plan, gated section by section → product-sunset-workshop

"I need to lose the legacy product without losing the legacy customer"

PMs sweat pricing, P&Ls, business cases, and build/buy/partner decisions. EOL rarely makes that list -- which is exactly why it goes badly. Nobody practices the thing they think they'll figure out when it happens. These prompts cover the full EOL lifecycle, and every one of them right-sizes the effort so a minor feature deprecation gets a short punch list while a flagship retirement gets the full cross-functional playbook.

The Goldilocks rule: every prompt asks how complex the sunset is before generating output. Tier 1 (feature deprecation) gets light-touch artifacts. Tier 2 (commercial product) gets the standard cross-functional process. Tier 3 (revenue-critical, hardware, or regulated) gets everything. The goal is coverage proportional to risk, not maximum ceremony.

"I need evidence about the market, not another meeting"

The answers live on the public web, and the AI should do the fieldwork — with citations, labeled inference, and no invented facts. Go to /market-intelligence/:

• Map a market's players and whitespace → market-landscape-scan
• Research a competitor set, just-enough style → competitive-research-snapshot
• Watch competitors on a schedule; report only material shifts → competitive-intel-watch
• Mine reviews and forums for unmet needs → voice-of-customer-miner
• Size a market with citations, bottom-up → tam-sam-som-analysis
• Build an evidence-cited battle card → battle-card-builder
• SWOT a company with sources, not vibes → swot-analysis
• Read an industry's structure and power → porters-five-forces
• Map growth options with evidence per quadrant → ansoff-matrix
• Run a collection sweep by intel discipline (OSINT, FININT, GEOINT/DEMOINT, TECHINT, HUMINT, SIGINT, MASINT) → the collection floor
• Fuse it all into confidence-rated intelligence → all-source-fusion
• Schedule the whole thing on a weekly/monthly/quarterly rhythm → fusion-cadence-routine

These are built to run unattended — in agent workflows, loops, and scheduled runs — with question budgets, materiality bars, and stop conditions.

The tradecraft behind them — the eight intelligence-community collection disciplines, signal → inference chains, confidence stacking, and regional source overlays — lives on the reference shelf: start with the Competitive Research Compendium.

"I need people to get it"

Evidence doesn't persuade; stories do. Go to /storytelling/:

• Turn research into a stakeholder narrative → Research-to-Narrative Bridge
• Visual storyboards, hero's journeys, Starts-with-Why arcs → browse the directory

"I run prompts as loops, goals, batches, or routines"

You still have to prompt the loop — nobody skipped that step, they just put a slash in front of it. Go to /loops/ for seasoned recipes at three levels (plain commands → plain-English controls → Just Enough Jinja2):

• Split epics until nothing splits, with a pass ceiling → story-splitting-loop
• Draft a PRD section by section, in dependency order → prd-section-loop
• Batch a backlog of epics without contaminating it → epic-story-batch
• Synthesize research one category at a time → research-synthesis-loop
• Run a versioned competitive watch that won't drift → competitive-watch-routine
• SWOT a competitor set with receipts, not vibes → swot-batch
• Size a market one segment at a time → market-sizing-loop

"I want to build my own prompts — or my own agents"

• Learn the patterns by using the meta-prompts in /prompt-generators/ — each one interviews you, then emits a reusable prompt
• Design a custom research agent → research-agent-prompt-generator
• Study the pattern documents (below), then the exemplars in /vibes/ for loop-safe, agent-ready prompt structures
• Reverse-engineer how any prompt works → /skeletons/
• Package a prompt as an agent skill your assistant invokes on its own → /skills/, where each folder holds a SKILL.md, an output template.md, and a worked example

"I want my assistant to reach for this without being asked"

/skills/ — Agent Skills: the same PM work packaged so a coding or chat agent loads it by itself when the moment matches. Each skill is one folder: a SKILL.md with YAML frontmatter the agent reads to decide relevance, the usual teaching comment block underneath, plus supporting template.md and examples/.

• Name the creatures stalking your product practice → dangerous-animals-of-pm-generator

These follow the Agent Skills open standard rather than a convention of my own, so they install anywhere skills are supported — Claude Code, claude.ai, or the Skills API. Drop a folder into ~/.claude/skills/ (or symlink it there from your checkout) and the skill is available as /skill-name. scripts/validate-prompts.py enforces the spec's constraints, so a skill that passes CI is a skill that will actually upload.

"I need to laugh so I don't cry"

/resumes-resignations-reactions/ — therapeutic workplace satire. You've earned it.

Can't find it? The generated catalog lists every asset with a one-line description.

---

How These Prompts Work (and Teach)

Three interaction modes power everything here — documented in interaction-modes.md:

1. Facilitation (you hold the context): the AI asks a budgeted 3–5 questions, one at a time, offering 3 context-aware recommendations plus "Other" — with standing escape routes ("take your best guess", or paste your notes to skip ahead). Spec: generative-guidance-pattern.md.
2. Checkpointed co-construction (an artifact holds the structure): the AI builds section by section against your template, stopping at a gate after each; gaps are labeled Assumption or Open Question, never invented.
3. Autonomous investigation (the world holds the context): the AI researches under an evidence contract — search plan shown first, real URLs, Fact / Inference / Assumption labels, and defaults that let it proceed when nobody's answering.

Prompts meant to run repeatedly — under loop or goal commands, or inside agents — use Jinja2 notation for explicit, bounded control flow: jinja2-prompt-structures.md.

The teaching layer: view the raw code

Every prompt carries an HTML comment block (<!-- ... -->) with its Description, Usage Note, When NOT to Use, and Pedagogic Notes — the instructor's notes explaining why the prompt works. GitHub's pretty preview hides them, so click Raw and copy everything, comments included. The AI ignores them; you shouldn't.

From user to builder to teacher

1. Use the prompts on real work
2. Read the comments to see why they're shaped that way
3. Customize for your industry and team
4. Build your own with the generators and pattern docs
5. Contribute back — see SUBMISSIONS-GUIDE.md

---

For Engineers and Agents Improving This Repo

• CLAUDE.md / AGENTS.md — the authoring contract, directory intent, and quality bar for AI-assisted sessions
• prompting-style-guide.md — the full methodology
• scripts/validate-prompts.py — structural checks (required metadata, v2 fixtures, companion-link resolution); errors block, warnings are the migrate-on-touch worklist
• scripts/generate-catalog.py — regenerates catalog/ from the metadata comment blocks (the blocks are the source of truth; the catalog is derived output)
• CHANGELOG.md — what changed in each version and what's deliberately left open

Run both scripts after any prompt edit. Existing v1-pattern prompts are grandfathered: migrate them to v2 when you touch them, and never mass-rewrite.

---

Contributing

Real problems, tested prompts, teaching comments. Start with SUBMISSIONS-GUIDE.md, then:

git clone https://github.com/deanpeters/product-manager-prompts.git
git checkout -b your-improvement-name
# make your changes, then:
python3 scripts/validate-prompts.py
python3 scripts/generate-catalog.py
# commit, push, open a pull request

Test across ChatGPT, Claude, and Gemini. Humans make the strategic decisions; AI assists.

---

Resources and Contact

• Productside Workshops — live training to go deeper
• Questions? Open an issue
• Connect: Dean Peters on LinkedIn

License

Licensed under CC BY-NC-SA 4.0 — use, adapt, and share freely for non-commercial purposes with attribution and share-alike. Commercial use requires expressed written permission from Dean Peters. Plain-English guide: LICENSING.md. (Versions prior to v2.3.0 were MIT; copies obtained under MIT remain MIT.)

Created by Dean Peters • Built for the product management community • Designed to teach through doing