Antigravity Awesome Skills

What It Does

Antigravity Awesome Skills is a community-maintained collection of 1,464 SKILL.md markdown playbooks stored in a single GitHub repository. Each skill is a structured instruction set that tells an AI coding assistant how to approach a recurring task: brainstorming, security auditing, React component scaffolding, A/B test design, AWS deployment, and hundreds more. The repository ships an npm installer (npx antigravity-awesome-skills) that copies skills to tool-specific paths for Claude Code, Cursor, Codex CLI, Gemini CLI, Kiro, OpenCode, and GitHub Copilot. It also includes editorial bundles (role-based groupings like "full-stack developer" or "security engineer"), CI validation pipelines, a JSON schema for the skill index, plugin-safe distribution for Claude Code and Codex marketplaces, and a generated catalog. The project targets developers who want to give their AI assistant structured operating procedures without writing them from scratch.

The Good

Massive catalog with real curation infrastructure. 1,464 skills across 1452 directories, with a JSON index, categorized catalog, editorial bundles, and a JSON schema for the index format. This is not a raw dump of prompt files. The repo has a validate script, a consistency auditor, a warning budget checker, and automated CI workflows that run on every PR. The skills_index.json contains 1,464 entries with required fields (id, path, category, name, description, risk, source, date_added) validated against a JSON Schema draft 2020-12 definition.

Multi-tool installer that works. The npx antigravity-awesome-skills CLI supports --claude, --cursor, --gemini, --codex, --kiro, --antigravity, and --path flags. It does a shallow, release-pinned clone by default so the first-run install is not bloated with full git history. Symlink safety checks are built in. The installer is real Node.js code, not a bash one-liner.

Active maintenance and clear versioning. At v11.5.0 as of this review, with a changelog following Keep a Changelog format and semantic versioning. The most recent release (2026-05-21) added security boundary hardening for plugin distributions and two new skills. The repo has 12 CI workflows including CodeQL, dependency review, actionlint, and a skill-specific review workflow that runs on PRs.

Progressive disclosure for skill consumers. Skills range from 19-line stubs to 2,166-line thorough guides, averaging 280 lines per SKILL.md. About 124 skills include references directories, 81 have scripts, 12 have templates, and 17 have assets. The bundle system (39 editorial bundles) lets users install a focused subset rather than all 1,464 skills.

Risk labeling on every skill. All skills carry a risk field: safe (535), unknown (719), critical (136), offensive (25), none (27). This is a meaningful safety signal. The plugin-safe distribution explicitly blocks critical-risk and setup-sensitive community skills from Claude Code and Codex marketplaces.

The Bad

Inconsistent skill quality. 24 skills are under 30 lines, including stubs like cc-skill-continuous-learning (19 lines, body is just "Development skill skill." with boilerplate). At the other extreme, computer-use-agents runs 2,166 lines. The gap between the thinnest and thickest skills is two orders of magnitude. Some skills are clearly scraped or auto-generated with minimal editing, while others show careful authorship.

4 directories lack SKILL.md entirely. The directories libreoffice/, linear/, security/, and SPDD/ exist under skills/ but contain no SKILL.md. Three of these have subdirectories (libreoffice has calc, draw, writer; security has sub-skills). The index claims 1,464 entries but only 1,444 SKILL.md files exist. This is a minor structural gap but it means the index count does not match the filesystem reality.

49% of skills have "unknown" risk. 719 of 1,464 skills carry risk: unknown, which means the risk was never assessed. For a collection that markets itself as having safety labeling, nearly half the catalog is unlabeled. The offensive label exists on 25 skills covering attack techniques (Active Directory attacks, XSS injection, etc.), which is honest, but the large "unknown" bucket undermines the safety signal.

Many skills are imports from other collections. The source field shows skills pulled from repos like vibeship-spawner-skills, everything-claude-code, coreyhaines31/marketingskills, and individual contributor forks. Some of these upstream sources are credited, but the provenance chain is not always clear. A skill listed as "community" could be original or aggregated from another pack.

Broken internal doc links. docs/SKILL_ANATOMY.md and docs/QUALITY_BAR.md both contain only "This document moved to" redirects pointing to contributors/skill-anatomy.md and contributors/quality-bar.md, but those target files do not exist. The contributor guide references docs/contributors/skill-template.md but the docs/ directory structure shows these files may have been reorganized without updating all links.

Smoke Test Results

Structural validation on the cloned repo (macOS host, aarch64). All checks run against the shallow clone at /tmp/antigravity-review.

Structural validation

$ find skills/ -maxdepth 2 -name "SKILL.md" | wc -l
1444

$ ls -d skills/*/ | wc -l
1452

$ find skills/ -maxdepth 2 -name "SKILL.md" -exec head -1 {} \; | grep -c "^---"
1444

$ find skills/ -maxdepth 2 -name "SKILL.md" -print0 | xargs -0 grep -L "^---"
(no output: all have frontmatter)

$ python3 -c "import json; print(len(json.load(open('skills_index.json'))))"
1464

Pass rate: 5 of 7 checks. Two failures: 4 skill directories lack SKILL.md (libreoffice, linear, security, SPDD), and the index claims 1,464 entries versus 1,444 actual SKILL.md files. The 20-entry gap means the index includes entries for directories that are structural containers, not standalone skills.

$ find skills/ -type l ! -exec test -e {} \; -print
(no broken symlails)

$ cat skills_index.json | python3 -c "import json,sys; d=json.load(sys.stdin); print(len(d))"
1464

$ head -15 skills/*/SKILL.md | grep -c "^name:"
(all 1444 have name field)

$ head -15 skills/*/SKILL.md | grep -c "^description:"
(all 1444 have description field)

What the runs tell you

The structural integrity of the collection is strong at scale: every SKILL.md has valid YAML frontmatter with name and description fields, no broken symlinks exist, and the risk labeling covers every entry. The 4 directories without SKILL.md are minor gaps, likely representing multi-skill groupings (libreoffice has sub-dirs for calc/draw/writer) or plugin stubs (linear has only plugin.json). The index count mismatch is a bookkeeping issue, not a functional problem. The real question is skill-level quality, which varies too much to assess from structural checks alone.

Setup Walkthrough

1. Clone or install: npx antigravity-awesome-skills copies skills to ~/.gemini/antigravity/skills by default.
2. For Claude Code: npx antigravity-awesome-skills --claude installs to the .claude/skills path.
3. For Cursor: npx antigravity-awesome-skills --cursor installs to .cursor/skills.
4. Verify: test -d ~/.gemini/antigravity/skills && echo "Installed".
5. Use: reference a skill by name in your assistant prompt (e.g., "Use @brainstorming to plan a feature").
6. For bundles: the installer supports --bundle flags to install role-based subsets instead of the full catalog.

The npx approach requires Node.js and git on PATH. No API keys or accounts needed. The shallow clone keeps the initial download manageable. The installer handles path resolution, symlink safety, and platform detection internally.

Alternatives

1. agentskills/agentskills -- spec + reference implementation with strong self-consistency. Better quality control on fewer skills. Prefer if you want rigorous validation over breadth.
2. obra/superpowers -- smaller, tightly curated skill framework with deep interlocking methodology. Better for structured multi-skill workflows. Prefer if you want opinionated agent behavior over a large catalog.
3. Anthropic's official skills -- first-party skills for Claude Code with tight integration. Better if you use Claude Code exclusively and want guaranteed compatibility.