An evaluated directory
of agent skills.

What this is

The agent-skill ecosystem is growing fast and is mostly impossible to navigate. Anthropic's official catalog has ~17 skills. OpenAI's has ~44. Then there's skillsmp.com (~545), skills.sh (~323), and the long tail on GitHub where anyone can drop a SKILL.md into a folder. None of those tell you which ones are actually well-made or worth adopting.

humangarden.ai is the layer on top. We scrape every source we can reach, score each entry on two independent axes — Craft (is the skill specified well?) and Adoption (is it maintained, documented, licensed permissively, used by others?) — and recommend the ones that pass both bars. Today we track 2,510 entries from 150 builders and publish 329 after curation.

This site does not redistribute or host any skill content. Every card links back to its source repository. We exist to evaluate, sort, and recommend; the canonical source remains the author.

Methodology

Every entry passes through the same pipeline: scrape → score → curate → publish. The score is the load-bearing part — it's how we decide which 249 of 1,291 entries make the cut.

Craft — D1 to D5

Craft scores how well the SKILL.md artifact itself is written. An LLM applies a five-dimension rubric and the dimensions are weighted into a 0–10 internal score (rendered as 0–5 in the UI).

1. D1 · Trigger clarity (25%)When should this skill fire? Are the conditions explicit and falsifiable?
2. D2 · Output specificity (20%)What does the skill produce? Is the output named and verifiable, not just "helpful answers"?
3. D3 · Scope precision (20%)How narrow is the skill? A skill that tries to do everything does nothing well.
4. D4 · Self-containment (20%)Does the skill bring its own dependencies, examples, and edge cases? Or does it assume context it doesn't have?
5. D5 · Reusability (15%)Does this work across projects and agents, or is it bolted to one codebase?

Adoption — A1 to A5

Adoption scores the practical reality of using the skill in production. No LLM involved — it's pure derivation from GitHub API data (last commit, license, README depth, stars, forks) plus our own knowledge of the author.

1. A1 · Maintenance (25%)Last commit recency, has releases, not archived. Active beats clever.
2. A2 · Documentation (25%)README length and structure. Are examples present? A demo?
3. A3 · License (15%)MIT/Apache/BSD permissive = full marks. Strong copyleft = half. None = zero.
4. A4 · Traction (20%)Stars and forks on a log scale. We don't worship star counts but they signal that someone other than the author has used it.
5. A5 · Authorship (15%)Do we have an editorial profile for this builder? What's the rest of their work scoring? A skill by a known builder has different priors than one by a stranger.

Composite — the headline

The score you see on every card is the composite, computed as the harmonic mean of Craft and Adoption. Harmonic mean (not arithmetic) means a strong axis can't compensate for a weak one — a beautifully-written SKILL.md that hasn't been touched in a year, or a popular repo with a hand-wavy spec, both drop below their best axis. The composite is honest about the worst.

Two formats: skill_md vs repo_tool

Not everything we index is a SKILL.md folder. When we profile a builder, we also pull their top 10 GitHub repos and index those as repo_tool entries — useful to know about, but they're not SKILL.md-format skills. repo_tool entries get only the Adoption score; their Craft column says "N/A — not a SKILL.md skill." We'd rather be honest than over-claim.

Publishing threshold

We publish when the relevant gate clears 4.0 / 5:

• skill_md entries publish when Craft ≥ 4.0 (legacy threshold preserved while we backfill Adoption metadata for older entries — see the public companion repo for the rolling list).
• repo_tool entries publish when Composite (= Adoption) ≥ 4.0.
• Below threshold, an entry stays in the directory (you can find it via search) but isn't recommended on the homepage or category pages.

Sources we scrape

Authority tier governs how much we trust a source's curation when there's a conflict. Tier 1 sources reflect the format owners' own picks; Tier 2 are curated marketplaces; Tier 3 is wild GitHub.

• Tier 1 (official): anthropics/skills, openai/skills
• Tier 1 (editorial seeds): Top 10 owner repos for every profiled builder
• Tier 2 (curated marketplaces): skillsmp.com, skills.sh
• Tier 3 (wild): GitHub code search for files named SKILL.md, plus repositories with a top-level marketplace.json

The scrape runs every 24 hours at 03:00 UTC. A new public skill that publishes today typically appears here within a day.

Editorial layer

Scores are auto-generated. The editorial layer is where humans (currently one human) take over and argue.

Tiers of depth

• Tier 1 — auto-indexed, score + summary, no editorial. The default for most entries.
• Tier 2 — reviewed. A short evaluation with what we tried and what we'd improve. Promoted manually from Tier 1 candidates.
• Tier 3 — deep-dive. A full thesis on why this skill matters, with worked examples. Rare; reserved for genuinely category-defining work.

Builder profiles

Five-block editorial reads of specific builders we track closely: their work, the thesis behind it, one transferable idea worth stealing. Each profile is grounded in observable output and links out with the builder's own framing where possible. We use @zarazhangrui as the gold standard against which every subsequent profile is measured.

Journal

Where we make claims that don't fit in a skill page — patterns we notice across the directory, methodology arguments, what the format owners' catalogs miss. One thesis post per week is the cadence we're aiming for.

Cadence + transparency

• Scrape: 03:00 UTC daily across 8 sources.
• Adoption refresh: in the same run, mechanical re-scoring from current GitHub state.
• Public source list: the awesome-agent-skills repo re-syncs at 04:00 UTC with the current top-50 picks. Star it if you want a heads-up via GitHub trending.
• Open audit trail: the source repo includes our entire scoring history in data/db.sqlite; you can clone and reproduce every number.

Contact + corrections

Found a wrong score, a stale source, a builder we should profile? Or you ARE a builder and we got your work wrong? Open an issue at homosapien1218/Awesome-agent-skills/issues. Corrections land in the next daily run.

Disclosure

We do not accept payment to rank, list, or feature any skill or builder. We hold no equity in the companies behind the skill marketplaces we scrape, and have no contractual relationship with Anthropic, OpenAI, or any source. Every score is generated by the pipeline described above. If we get something wrong, the recourse is a GitHub issue and a corrected score in the next 24 hours — not a paywall.