PMX-Skills

Product Management AI Skills

Skills

Skill	Purpose
first-principle-thinking	Challenge assumptions, identify fundamentals, and build a testable recommendation.
ghost-writer	Learn a user's voice from their own writing, draft in context, and improve from corrections.
weekly-prioritization	Reconcile goals, evidence, capacity, and active work into priority decisions.
feedback-triage	Find customer needs and traceable themes, then propose proportionate actions.
okr-coach	Improve objectives and results, review measured progress, and gather confidence.
business-review	Connect outcomes, learning, risks, and next-period decisions in a review or deck outline.
team-context-update	Share relevant behind-the-scenes context without repeating the delivery board.
knowledge-base-curator	Bootstrap or maintain a source-backed Markdown wiki.
visualization-critique	Review charts for accuracy, clarity, accessibility, and evidential limits.
product-planning-interview	Elicit product needs, boundaries, risks, and success criteria before planning.
html-artifacts	Build portable HTML reports, comparisons, decks, maps, and local interactive artifacts.
pmx-canvas	Operate a PMX Canvas workbench through MCP, HTTP, or CLI, with persistent nodes, edges, and context pins.

The skills live in .agents/skills/ for project discovery. To use them elsewhere, copy the desired skill directory, including its references, into that project's .agents/skills/ directory. In a compatible client, invoke /first-principle-thinking or /ghost-writer, or ask for the described task naturally.

The generic workflow skills work independently from supplied context with optional integrations. pmx-canvas is a runtime-specific exception: it requires a compatible PMX Canvas installation. No skill requires a private repository or organization-specific service. Examples and evaluation scenarios are synthetic. Keep personal profiles, customer records, and actual business knowledge in an appropriate user-controlled workspace, not in this shared skill collection.

Agent Plugin

The repository root is also a skills-only Agent Plugins 1.0.0 package named pmx-skills. Its plugin.json manifest targets the portable standard, not an Amp-specific TypeScript plugin API.

plugin.json
skills -> .agents/skills
.agents/skills/
└── <skill-name>/
    ├── SKILL.md
    ├── evals/
    └── references/  (where needed)

Load the repository root through an Agent Plugins-compatible client's local-directory plugin installation flow. The client must support the skills component type; installation commands, namespacing, and enablement are client-specific and not defined by the standard. Do not enable both the plugin and separately installed copies of the same skills unless the client handles duplicates.

The skills directory is a relative symbolic link to the existing .agents/skills source of truth. All targets and bundled resources stay inside the plugin root, as required by the specification. Edit skills in .agents/skills/; no generation or synchronization step is needed, and existing project skill discovery remains unchanged.

Clone or copy the complete package, including .agents/skills, with symlinks preserved. On systems or packaging tools without symlink support, materialize skills/ as a real copy of .agents/skills/ in a separate distribution directory, with plugin.json at that directory's root. Do not replace the tracked link with a second maintained copy. Copying only the manifest and the link leaves an incomplete package.

This plugin has no MCP servers, hooks, executable entry point, credentials, or client extensions. Loading the plugin does not authorize publishing files or accessing private accounts.

Examples

• First principles: "We think we need a new onboarding tool. Challenge that assumption and suggest the smallest experiment that would tell us what to do."
• Ghost writing: "Here are three emails I wrote. Learn my voice, then draft a reply using the facts below."
• Recalibration: "This sounds too formal. Here is how I would write it. Update my email style to reflect this."

ghost-writer starts by requesting user-authored samples. It adapts within the conversation and can save a voice profile at a user-approved private location when asked. Personal samples and profiles do not belong in this shared repository. No personal identity, language, employer, or writing style is preconfigured.

Starting a wiki

Use PMX Context Frame Example as the reference knowledge-base structure: maintained knowledge in wiki/, source evidence in raw/, unprocessed intake in inbox/, and generated outputs in artifacts/. Its wiki includes navigation, a dated change log, product context, decisions, and testable assumptions. It is a template with placeholders, not a populated product dataset. The skill describes and links to it; its files are not bundled here.

knowledge-base-curator includes its own bootstrap guidance. It first looks for an existing knowledge base in the selected workspace and reuses its conventions. If none exists, it confirms the destination and privacy expectations, then creates a minimal structure such as:

wiki/
├── index.md
└── pilot-decision.md

Topic pages are created from actual source material, not empty templates. A sources/ folder is optional and created only when retaining local source copies is authorized. No Obsidian, QMD, database, or pre-existing wiki is needed. The skill preserves sources, links claims to evidence, and avoids duplicate pages on repeated ingestion. Without file-writing access, it returns copyable Markdown instead of claiming it saved a wiki.

Grow toward the reference structure only as needed. If the workspace already uses raw/ for evidence, reuse it instead of adding a duplicate sources/ folder. Neither the example's agent runbook nor its full folder hierarchy is required.

HTML artifacts

html-artifacts combines a single-file authoring workflow with a portable catalog of 19 PMX Canvas-inspired primitive types. It maps comparisons, timelines, maps, reports, decks, galleries, and local editors to native HTML/CSS/SVG—not Canvas runtime APIs. No specific branding, CDN, framework, or backend is required.

Example: "Turn these rollout options into an offline HTML decision one-pager with a comparison grid and milestone timeline. Check desktop, mobile, and print layouts. Do not publish it."

The skill includes guidance for evidence integrity, safe text rendering, keyboard interaction, print/no-JavaScript fallbacks, and actual export verification. Browser rendering checks apply to generated artifacts; the bundled evaluation prompts are scenarios to run, not pre-rendered templates or a claim that all primitive types have been browser-tested.

PMX Canvas

pmx-canvas is imported from the upstream product with its complete reference set and evaluation fixtures. It covers persistent spatial boards, workspace identity checks, nodes and edges, context pins, live schemas, sessions, and human steering. Use html-artifacts for standalone HTML files; use pmx-canvas when working in the actual Canvas workbench.

The skill does not install or start the runtime or configure MCP automatically. Consult its installation reference when needed, and check the installed runtime's version and live schemas before using version-specific commands. Run runtime evaluations only against a disposable Canvas workspace.

The upstream record identifies the imported revision and local adaptations. The original MIT license is bundled with the skill so its required attribution survives copying. Public project references and copyright attribution are retained; personal profiles and employer details are not imported.

Design and checks

Created using Anthropic's skill-creator guidance. ghost-writer generalizes an existing personal ghost skill's source-grounding, register selection, drafting, and editing workflow; its personal corpus and preset persona are intentionally excluded.

Each skill includes representative prompts and expected outcomes in evals/evals.json. These are regression scenarios for future skill runs, not a claim of benchmarked voice quality. Voice fidelity requires feedback from the person whose writing is being modeled.

Feedback triage includes a runnable synthetic source double and behavioral run guide for pagination, full-context reads, access failures, and supplied-only/no-tool boundaries. It runs through Python and records call traces; it does not connect real accounts or test an MCP transport. Canvas's installation reference describes its separate live-runtime smoke check. Neither replaces running the remaining skill scenarios in representative host environments.

Run file-writing evaluation scenarios only in disposable workspaces. For scenarios describing an existing wiki, provide synthetic index/topic/source fixtures matching the prompt before running the skill. Never use the skills repository or real private notes as an evaluation destination.

To check the evaluation JSON syntax from the repository root:

for file in .agents/skills/*/evals/evals.json; do
  python3 -m json.tool "$file" > /dev/null || exit 1
done

This checks JSON syntax, not skill behavior or writing quality.

License

The original collection is licensed under MIT. The vendored pmx-canvas skill retains its upstream MIT notice. Preserve applicable notices when redistributing the collection or copying skills.