The architecture is intentionally boring. A single SKILL.md file acts as a router: it maps 40+ user intents to 6 knowledge modules loaded on demand from a `knowledge/` directory. Each module contains frameworks, decision trees, quality gates, and anti-pattern detectors. There are 12 templates in a `templates/` directory. Total system: ~130KB across 20 Markdown files. No build step, no runtime, no API calls. You can read every line before installing.
What makes it different from just prompting "act like a PM": the knowledge modules encode actual domain logic. The finance module has 32 SaaS metrics with exact formulas and stage-specific benchmarks. Ask it about 8% monthly churn and it doesn't say "that's high" -- it computes `1 - (1 - 0.08)^12 = 63%` annual churn, flags it as a red severity indicator, and tells you the benchmark for your stage. The discovery module enforces Mom Test principles and catches "Solution Smuggling" (sneaking a solution into a problem statement). The artifacts module has 8 story splitting patterns and 9 epic breakdown patterns, not just a template to fill in.
The quality gate system is the part I'm most interested in feedback on. Every output passes through universal gates (assumptions must be labeled, outcomes must be measurable, no "Feature Factory" outputs) plus domain-specific gates loaded from the active knowledge module. The agent is designed to push back -- if you ask for a PRD without a clear problem statement, it will challenge you rather than generate filler.
Tradeoffs worth noting: it's CC BY-NC-SA 4.0, not MIT, since the knowledge modules represent significant curation work. It's also opinionated by design, which means it won't work well if you want a yes-machine. And it's optimized for Claude Code's SKILL.md convention -- it works in other editors that support Markdown skills but the routing system is tuned for Claude's context window management.
GitHub: https://github.com/Digidai/product-manager-skills
Install: `clawhub install product-manager-skills`
