magnus919/agent-skills

What it claims

A single-author catalog (the README speaks in first person: "where I codify practical patterns") of Agent Skills. The summary does not say how the skills are built, and the skills are built in three recurring ways:

• A uniform agent-facing CLI contract. Many integration skills ship a "bounded stdlib CLI" with --json, --limit, and a --dry-run/--yes gate on any mutation. This applies to HubSpot, Notion, Slack, Stripe, SendGrid and Terraform (tfops, which adds --force). Several ops skills ship read-only probes instead: pgdiag for PostgreSQL, litellm-health, vllm-health, telemetry-check, k8s-cli, and pwrun for Playwright.
• Explicit routing between skills. Most descriptions carry "Do not use for…" clauses and name the sibling skill that owns adjacent work. For example, ai-governance routes legal questions to legal-strategy. Thin "orchestration" skills (product lifecycle, the production gate model, the delivery operating model) compose specialists rather than duplicate them. The skills are therefore designed to be installed together, not cherry-picked.
• Shipped evals. Many skills state an eval count, typically 5–10 output-quality or safety scenarios per skill, alongside templates and "dated references".

How to use it

The only concrete install command on the page is for the debate tool, agent-council. It is a pip package built on PydanticAI and PydanticGraph and needs one model API key:

pip install pydantic-ai && pip install agent-council

The page links "Install the full catalog" and "Keep skills up to date", but the fetched text does not contain those commands.

A meta-skill documents the format itself: directory structure, frontmatter schema, naming, and progressive disclosure. The repo tells you to load it when authoring or reviewing skills.

The build-cli-for-agents skill (exact name not shown on the page) codifies 10 patterns behind the CLI contract above:

• non-interactive
• --json
• --dry-run
• idempotent
• lazy auth
• progressive help

It also ships an agent-compatibility test suite and a bash scaffold. The author says these patterns are drawn from failures building 15+ agent-facing CLIs. This is the most reusable piece if you are writing your own tools.

Limits

• No licence on the page. The fetched text states none. The only licensing note is that the systematic-debugging skill is "Adapted from obra/superpowers (MIT)". Check the repo's LICENSE before reuse.
• Mixed scope. Not everything is engineering. The catalog includes a DSM-5-TR mental-health assessment skill, scoped as "not for formal diagnosis, treatment decisions, crisis intervention". It also has nonclinical life coaching, an occult/CRT-noir image aesthetic, a Nous Research branding skill, and Raleigh-specific open-data and ASCII-city packs. Installing the full catalog installs all of these.
• Overlap and gaps. Two near-duplicate FlareSolverr skills exist; FlareSolverr is used to get past browser challenges, which may conflict with target sites' terms. Many integration CLIs cover only a narrow slice of the API. For example, HubSpot supports contacts, deals and pipelines, and Stripe cancellation is the only mutation.
• Unverified numbers. Figures quoted inside skills come from the skill text, not from evidence on this page. The LangGraph skill's "~94% accuracy" for the supervisor pattern and "~40% fewer LLM calls" for swarm have no stated source or setting.
• Incomplete text. The fetched page was cut at 60,000 of 71,286 characters (not 40,000, as the summary says). The list stops mid-entry at a waitlist-portal skill, so later skills are not reflected here.