pm-claude-skills: 1153 Agent Skills as Plain Markdown Files

Why a folder of markdown files is the whole product

The problem PM Skills targets is not missing capability in a model. It is that a general assistant answers open-ended professional questions with the safe, average answer. The README makes this argument directly: LLMs are "too correct", and on open-ended questions they return something technically right and forgettable. A skill is an attempt to replace that default with a named framework, an output template, quality checks and a list of anti-patterns, so the model follows a structure instead of improvising one.

The audience is wider than the name suggests. PM here stands for Professional, not Product Management, and the repository says so in the README. The catalogue covers work tasks (write a PRD, run the AGM, decode a lease) and personal ones (a medical bill that makes no sense, a layoff, a death in the family). The Skill Packs entry point is organised by life moment rather than by job title, which tells you who the maintainer thinks is arriving: someone with a specific problem, not someone browsing a taxonomy.

The distribution model follows from that. A skill is a markdown file, so there is nothing to execute, nothing to sandbox and nothing to monitor. That is a real advantage over plugin systems that ship code, and it is also the ceiling on what this project can do.

What a SKILL.md actually contains

Each skill is a single SKILL.md file. The README describes the contents as the framework, an output template, quality checks and anti-patterns. The repository points readers at skills/lease-decoder/SKILL.md as a representative example and states that the files are designed to be read by humans too, which is an unusual thing to claim and an easy one to verify.

Around the skills sit several other artifact types. The package.json files list includes agents/, commands/, output-styles/, exports/, templates/ and connectors/ alongside skills/, so the shipped surface is broader than the skill files alone. The README mentions subagents and slash commands, and the repository has a .claude-plugin/ directory, which is the layout Claude Code expects for a plugin. The README says the project is in Anthropic's official plugin directory and that a Claude Code user can type /plugin, search pm-skills and install.

For every other tool, the mechanism is export rather than integration. The README describes ready-to-paste exports for ChatGPT, Gemini, Cursor, Codex and eight more tools, stored under exports/. There is no shared runtime between them; the same markdown is reformatted per target. That is why the language in package.json calls the package a CLI rather than a library.

Installing with npx pm-claude-skills add

The CLI is the primary install path for anything that is not Claude Code. The package description is explicit that you do not run npm install: the binary is exposed as pm-claude-skills and the intended invocation is npx with the add subcommand. Running it without arguments opens a menu.

npx pm-claude-skills add

The README says the menu lets you pick your tool: Cursor, Codex, Windsurf, ChatGPT, Gemini and others. The package.json description shows the non-interactive form with an --agent flag, which is what you would use in a script or a setup document.

npx pm-claude-skills add --agent <tool>

On Claude Code the path is different and shorter. The README gives it as a three-step sequence inside the tool itself: type /plugin, search for pm-skills, install. After that, the README's example prompt is a plain request in the conversation, not a command.

decode this lease

If you want to look before installing anything, the project runs a browser playground at its GitHub Pages site where you fill a short form and a skill runs without an account. The README states that your key stays in your browser. There is also a find page that takes a plain-language description such as "my landlord kept my deposit" and names the matching skill.

The Org Edition container and what it does not solve

The Dockerfile is short and its comments explain the intent: the Org Edition as a container, your company's skill library in one command. It builds from node:20-alpine, copies org/, web/, skills/ and workflows.json, exposes port 8080, and runs node org/server.mjs --port 8080 as the non-root node user. Private skills are mounted rather than baked in.

FROM node:20-alpine
WORKDIR /app
COPY org/ ./org/
COPY web/ ./web/
COPY skills/ ./skills/
COPY workflows.json package.json ./
EXPOSE 8080
USER node
CMD ["node", "org/server.mjs", "--port", "8080"]

The limitation is structural rather than a bug. A skill is text that instructs a model. Nothing in the container enforces that the model followed the framework, and nothing validates the output against the quality checks that ship inside the SKILL.md. The repository does carry conformance/REGISTRY.md and GitHub workflows named skillcheck and skill-audit, so there is tooling around skill quality, but the README does not describe a runtime gate on generated output. If your requirement is that every PRD leaving the building conforms to a schema, this library is a prompt layer, not a validator.

The second limitation is scope. 1153 skills is a lot of surface area for one maintainer, and the release cadence in the repository is fast: v76.2.0 and v76.2.1 both landed on 2026-08-16, and v77.0.0 on 2026-08-18. The last push was on 2026-08-18. Frequent releases mean frequent re-installs if you pin nothing, and the README does not document a version pinning or rollback procedure for installed skills.

How it compares with building your own skill folder

The obvious alternative is writing your own SKILL.md files. The repository is candid about this being possible: it ships SKILL-AUTHORING-STANDARD.md, SKILLSPEC.md, scripts/new-skill.mjs and scripts/new-bundle.mjs, and CONTRIBUTING.md. If you have strong opinions about how your team writes a postmortem, a hand-written skill encodes those opinions exactly and stays under your review.

The difference in approach is coverage against fit. A hand-written folder gives you ten skills that match your process and nothing for the lease you have to sign next week. PM Skills gives you breadth across work and personal tasks, with the trade-off that each framework reflects the maintainer's judgement rather than yours. The pm-thinking and pm-focus bundles illustrate the split: the first is built to attack your own plan (red-team-my-plan, poke-holes-in-this, devils-advocate-on-demand), the second to get you moving (where-do-i-start, task-to-first-step, the-one-thing). Those are opinionated behaviours, not neutral utilities, and you will either agree with the framing or find it grating.

A second alternative is doing nothing and prompting the model directly. That costs nothing and works for common tasks where the model's default is already good. The case for a skill is the task you do once a year, where you have no personal template and no way to tell a good output from a plausible one.

Licence, maintenance and upgrade cost

The licence is MIT, stated in package.json and present as a LICENSE file at the repository root. MIT is permissive, so copying, modifying and redistributing the skill files is allowed provided the copyright notice and permission notice are retained. That matters more here than for a typical dependency, because the artifact you are shipping into your AI tool is the text itself. If you fork a skill and edit it internally, the MIT terms travel with the file. This is a description of the licence, not legal advice; if you are redistributing the skills inside a commercial product, read the LICENSE file yourself.

The maintenance signal is the release history. v77.0.0 landed on 2026-08-18, two days after v76.2.1, and the repository is not archived. The upgrade cost is low in absolute terms because there is no dependency graph to resolve: an update replaces markdown files. The cost that is easy to miss is review. If your team has customised a skill, a re-install can overwrite that edit, and the README does not document how the CLI handles locally modified files. Check bin/cli.mjs before you let an automated update run against a customised library.

There is also a sponsorship link and a newsletter in the README. Neither changes the licence, but the project is maintained by one author with funding links, which is worth knowing when you decide how much of your process to encode in its files.

Editorial conclusion

Adopt it if you already use Claude Code, ChatGPT, Cursor or Codex and want a written framework for a task you do rarely and badly, such as decoding a lease or running a blameless postmortem. Do not adopt it if you expect the skills to execute anything: they are markdown instructions for a model, with no runtime and no telemetry, and the README states that installing copies text files and uninstalling is deleting them. Before you commit, open skills/lease-decoder/SKILL.md and read it end to end, because the README says the files are designed to be read by humans too, and that is the only way to judge whether a given framework matches how your team actually works.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes