pm-skills-arsenal

pm-skills-arsenal is a pack of 12 product-management agent skills by Parth Sangani (MIT, billed as personal research separate from his day job). Each skill is a single long SKILL.md — the README’s pitch is “codex-grade” skills of 1,000–1,300 lines each, aimed at senior PMs, founders without a dedicated CPO, and AI engineers wiring PM methodology into agents. It covers the same non-coding vertical as pm-skills (Paweł Huryn’s marketplace) but bets the opposite way on shape: a few very deep skills rather than 68 shallow ones across 9 plugins.

The 12 skills

Four groups spanning the PM lifecycle:

• Analysis & Research — Competitive & Market Analysis, Discovery Research, Problem Framing
• Definition & Measurement — Specification Writing, Metric Design & Experimentation
• Strategy & Planning — Product Strategy, Go-to-Market Strategy, Pricing & Packaging
• Communication & Influence — Executive Writing, Narrative Building, Multi-Channel Publishing, Stakeholder Alignment

Typed, machine-routable skills — the distinctive move

Where most skill packs lean on a prose description for the agent to match against, every SKILL.md here declares a typed capability_summary, input_schema, and output_schema in frontmatter. The stated point is that an orchestrator can route between skills structurally — picking and chaining them from their I/O contracts without reading the full body. That treats a skill less like a loaded document and more like a typed function: the open skills “advertise → load → read” progression pushed one rung further, from a metadata blurb to a declared input/output signature. It is the clearest instance the spoke has of skills reaching for composability-by-schema rather than composability-by-prose.

Distribution — four install paths, deliberately harness-agnostic

1. APM (Agent Package Manager) — one apm.yml manifest targeting many harnesses (GitHub Copilot CLI, Claude Code, Cursor, …); the multi-harness reach the standard anticipates, expressed as a package manifest.
2. Plugin marketplaces — Claude Code, the Agency marketplace (Microsoft Copilot), awesome-copilot.
3. MCP server — exposes five tools: list_skills, get_skill, list_agents, get_benchmark, run_skill. So the pack ships as a queryable service, not just files — an agent can enumerate and invoke a skill over MCP (the “skills × MCP = agency” formula, here with run_skill as the invocation surface).
4. Direct — copy a SKILL.md into any LLM context.

Why it matters here

A second PM-skills instance is not just a duplicate: together they mark a small design axis inside one vertical — many shallow skills chained by slash-commands (pm-skills) vs few deep skills routed by typed schemas (here). The typed-schema + run_skill-over-MCP framing is the sharper contribution: it’s the vertical-generalization thread (ai-job-search, designpowers, pm-skills) pushed toward structural, machine-negotiated composition — a candidate answer to how orchestrators pick skills at scale without reading each one.

Tier T2 / caveats. Primary source for what it is (structure, schemas, distribution), but the quality framing is self-marketing: “codex-grade” and the get_benchmark tool are the author’s own claims with no independent evaluation — the same READMEs-only evidence gap the spoke’s Open questions flag. Line-count is a proxy for depth, not a measure of it.

Cross-spoke context: the PM subject matter (discovery/strategy/GTM/pricing) is the content of the skills; the artifact is agent tooling. Faint seams to ../search-marketing-wiki (its GTM / pricing skills) and ../research-wiki (PM discovery as a tools-for-thought practice) — cross-link, don’t duplicate.

Related

pm-skills · agent-skills · agentskills-spec · model-context-protocol · agent-orchestration · agents-md · ai-job-search · designpowers · superpowers · parth-sangani