PM Skills Marketplace

What It Does

PM Skills Marketplace is a collection of 68 agent skills and 42 chained commands organized into 9 independently installable plugins. Each skill encodes a product management framework (Lean Canvas, Opportunity Solution Trees, PRD templates, North Star Metric, A/B test analysis) as a structured prompt that an AI agent follows step by step. Commands chain multiple skills into guided workflows: /discover runs a full discovery cycle from ideation through experiment design, /write-prd generates a product requirements document from a feature idea, and /ship-check turns a vibe-coded codebase into a reviewer-ready shipping packet.

The target users are product managers, startup founders, and AI-assisted developers who want structured PM workflows inside their coding agent rather than generic AI text output. It is designed for Claude Code and Claude Cowork, with documented compatibility for Gemini CLI, Cursor, Codex CLI, OpenCode, and Kiro.

The Good

110 components across 9 plugins, and the validator passes clean. The validate_plugins.py script checks every plugin.json manifest, every SKILL.md frontmatter (name matches directory, description present), every command file (description plus argument-hint), README presence, and cross-references. All 9 plugins pass with zero errors and zero warnings. That is 68 skills, 42 commands, and 9 manifests all validated in under a second. This is the cleanest structural validation result of any doc-only skill pack reviewed.

Each skill embeds a real PM framework, not generic advice. The Opportunity Solution Tree skill references Teresa Torres. The brainstorm-experiments skill references Alberto Savoia's pretotyping. The North Star Metric skill classifies business into Attention, Transaction, and Productivity games. The prioritization-frameworks skill covers 9 frameworks (Opportunity Score, ICE, RICE, MoSCoW, Kano, and others). These are not surface-level summaries. The OST skill runs 60 lines of structured instructions including the four-level tree, scoring formulas, and process steps. The create-prd skill defines an 8-section template with specific content requirements per section.

The AI Shipping Kit plugin is novel. The pm-ai-shipping plugin addresses a problem no other reviewed skill pack touches: how do you review AI-generated code that has no documentation of intent? The shipping-artifacts skill defines a core-plus-conditional documentation set (architecture, flows, permissions, variables, tests, plus optional emails/cron/SEO/automation docs). The intended-vs-implemented skill defines a method for finding the gap between documented intent and actual code behavior. The /ship-check command chains these into a full audit pipeline. This is the most thoughtful treatment of the "vibe coding" accountability problem I have seen in a skill pack.

Command chaining is well-designed. Commands use skills as building blocks. /discover chains four skills: brainstorm-ideas, identify-assumptions, prioritize-assumptions, brainstorm-experiments. Each command includes checkpoints where the agent pauses for user input before proceeding. Commands suggest next steps in natural language (never hard-referencing commands from other plugins, since plugins install independently). This composability means adding one command leverages multiple existing skills.

Progressive disclosure keeps loading lean. Frontmatter is kept short (name and description, plus trigger phrases for auto-loading). The body is loaded only when the skill is triggered. The CLAUDE.md guidance explicitly instructs contributors to keep frontmatter lean and put detail in the body. Word counts across the 68 SKILL.md files average roughly 600 words each, which is in the right range for structured agent instructions.

Multi-agent compatibility is documented concretely. The README includes copy-paste commands for Gemini CLI, OpenCode, Cursor, and Kiro. The Codex CLI install path is documented with the caveat that slash commands do not translate (Codex plugins do not expose commands), along with a workaround: ask Codex to convert command files into native skills. This is honest and practical.

The Bad

All 68 skills are doc-only. No executable scripts, no tests to run. This is the nature of the domain: PM frameworks are prompt templates, not code. But it means the smoke test for this review is structural validation and reading, not running real workflows through an agent. A skill like sql-queries (generate SQL from natural language) or ab-test-analysis (statistical significance calculations) could benefit from companion scripts that verify output quality. The AI Shipping Kit commands (/security-audit-static, /performance-audit-static) are described as "static audits" but are still just prompts. There are no actual static analysis scripts backing them.

Some skills are thin wrappers around common knowledge. The swot-analysis skill (SWOT with actionable recommendations) and pestle-analysis skill are 40-60 line summaries of frameworks that any competent LLM already knows. The value comes from the structured template and consistent formatting, not from domain knowledge the model lacks. Users who expect deep specialization in every skill will find some entries light.

No example outputs or before/after demonstrations. The README shows example prompts ("What are the riskiest assumptions for our AI writing assistant idea?") but no example outputs. For a skill pack targeting PMs who may not be deeply technical, showing a sample PRD, a sample OST, or a sample competitive analysis would help buyers evaluate whether the output quality meets their standards. The repo includes images of the plugin browser and installation flow, but no sample generated documents.

Cross-plugin discovery depends on the agent, not the marketplace. The "no cross-plugin references" design rule means there is no index or graph connecting related skills across plugins. If a user runs /write-prd and then wants to stress-test the result, they need to know about /red-team-prd in the same plugin, or /discover in a different one. The README lists all commands, but the agent will not automatically suggest cross-plugin workflows unless the user asks.

Smoke Test Results

Hands-on review on macOS arm64. Cloned the repo, ran the validator, inspected all 9 plugin manifests, read 12 representative skills across all plugins, and checked 6 command files for chaining correctness.

Run A: Clone and validate

$ git clone --depth 1 https://github.com/phuryn/pm-skills /tmp/pm-skills
Cloning into 'pm-skills'...

$ cd /tmp/pm-skills && python3 validate_plugins.py
Found 9 plugins

======================================================================
 Plugin Collection Validator. Report
======================================================================

┌─ pm-ai-shipping [2 skills, 5 commands] ✓ PASS
┌─ pm-data-analytics [3 skills, 3 commands] ✓ PASS
┌─ pm-execution [16 skills, 11 commands] ✓ PASS
┌─ pm-go-to-market [6 skills, 3 commands] ✓ PASS
┌─ pm-market-research [7 skills, 3 commands] ✓ PASS
┌─ pm-marketing-growth [5 skills, 2 commands] ✓ PASS
┌─ pm-product-discovery [13 skills, 5 commands] ✓ PASS
┌─ pm-product-strategy [12 skills, 5 commands] ✓ PASS
┌─ pm-toolkit [4 skills, 5 commands] ✓ PASS

======================================================================
 Summary
======================================================================
 Plugins: 9
 Skills: 68
 Commands: 42
 Total: 110 components

 ✓ ALL CHECKS PASSED (0 warnings)

PASS. All 110 components validated with zero errors and zero warnings.

Run B: Structural spot checks

$ find /tmp/pm-skills -name "SKILL.md" | wc -l
68

$ find /tmp/pm-skills -path "*/commands/*.md" | wc -l
42

$ find /tmp/pm-skills -name "plugin.json" | wc -l
9

$ wc -w $(find /tmp/pm-skills -name "SKILL.md" -exec cat {} +)
 40965

PASS. Counts match the README claims (68 skills, 42 commands, 9 plugins). Total skill content is approximately 41,000 words across 68 files, averaging 600 words per skill.

Run C: Frontmatter and name-matching check

$ for f in $(find /tmp/pm-skills -name "SKILL.md"); do
 dir=$(basename $(dirname "$f"))
 name=$(head -5 "$f" | grep "^name:" | sed 's/name: *//' | tr -d '"')
 if [ "$dir" != "$name" ]; then echo "MISMATCH: dir=$dir name=$name"; fi
 done
(no output)

PASS. Every skill's frontmatter name field matches its directory name.

What the runs tell you

The structural validation is airtight. All 110 components pass the validator with zero warnings, names match directories, counts match documentation, and the 9 plugin manifests are consistent. The limitation is that these are doc-only prompt templates: the only way to verify output quality is to run them through a real agent with a real PM scenario, which is beyond what a structural review can confirm. The validator gives high confidence that the plumbing is correct; the PM content quality is best evaluated by domain experts.

Setup Walkthrough

1. Clone or install via the Claude Code plugin marketplace: claude plugin marketplace add phuryn/pm-skills
2. Install individual plugins: claude plugin install pm-execution@pm-skills (repeat for each desired plugin)
3. Verify by invoking a command: /write-prd Smart notification system that reduces alert fatigue
4. For other agents (Gemini CLI, Cursor, OpenCode): copy skill folders to the appropriate directory as documented in the README.

No build step, no dependencies, no runtime requirements. The skills are markdown files that the agent reads on demand.

Alternatives

1. addyosmani/agent-skills -- a broader SDLC methodology pack (23 skills) that includes some PM-adjacent skills (planning, code review) but is focused on engineering workflow, not PM practice. Better for developer-centric teams. Less PM depth.
2. K-Dense-AI/scientific-agent-skills -- a research-oriented skill pack with literature review, experiment design, and hypothesis testing. Overlaps with the discovery and analytics plugins but targets academic research rather than product management.
3. anthropics/skills -- the official Claude Code skill collection, which includes general-purpose skills that overlap with some PM-toolkit entries (resume review, writing). Lighter on PM-specific frameworks and does not include the structured command chaining.