Open-source Agent Skills for Anthropic Claude, published by Neural Cube under Apache 2.0.
Each skill in this monorepo solves a real problem in working with Claude across long-running projects. Skills are designed to produce outputs that travel cleanly across AI surfaces (claude.ai Projects, ChatGPT Projects, Copilot M365 Projects) so the same context grounds future sessions wherever the operator works next.
| Skill | Version | Download | What it does |
|---|---|---|---|
| project-context | 0.7.1 | .skill |
Captures forward-grounding context from a conversation into a three-file project memory (active, entities, archive) that future chats read to start grounded. Versioned file identity, decoupled data schema, and a mandatory pre-flight check before any write. |
| session-handoff | 0.1.1 | .skill |
Captures a high-fidelity, machine-readable handoff of a working session that any agent on any platform can resume cold. Five modes (generate, update, consolidate, retrofit, share-sanitize) behind one auto-routed entry point. |
| deep-analysis | 0.2.0 | .skill |
Frontier-class deep-analysis: one maximum-extraction pass over an artifact (codebase, repo, document set, website, architecture, product), emitting dense machine-readable deliverables for execution-class consumers. Six lenses; default survey + review. Every recommendation survives a war-game pass. |
| conversation-recap | 0.1.1 | .skill |
Narrative "Previously On" recaps that re-immerse you in a stale thread rather than re-briefing you on it. 24 registers, three length tiers, and a fabrication rule that keeps the shaping honest. |
| session-orchestrator | 0.2.0 | .skill |
A three-layer session model for multi-issue software work, where the session that decides is never the session that writes the code. Carries ten verification rules, each paid for by a real failure. Hub-resident: see its README before installing. |
Each release also ships an identical .zip (some surfaces prefer it) and a MANIFEST.sha256.
Note on download links. GitHub's
releases/latestis repository-wide, and this repository tags per skill, so there is no per-skill "latest" URL. The links above are pinned to each skill's current release tag and are updated as part of every release.
Download the .skill file for the skill you want and add it to your Claude surface. The .skill file is a plain zip whose single top-level folder matches the skill's name, so if a surface wants a folder or a .zip instead, rename or unpack it.
Every skill is versioned independently on dotted semver, declared in the version field of its SKILL.md frontmatter. Directory names never carry a version. Tags are <skill-name>-v<MAJOR>.<MINOR>.<PATCH>.
Skills stay on 0.x until proven in real use; each skill's ROADMAP.md states its own conditions for 1.0.0.
skills/<skill-name>/ one directory per skill; the only place a SKILL.md lives
docs/<skill-name>/ development documents, per skill
scripts/ build and validation tooling
Contributors and agents: see CLAUDE.md for the full layout, naming, versioning, and release contract, and AGENTS.md if you are a non-Claude agent.
Apache 2.0. See LICENSE and NOTICE.
See CONTRIBUTING.md. Every skill folder must include SKILL.md, README.md, CHANGELOG.md, ROADMAP.md, and USAGE.md.
Active. New skills land via reviewed pull requests.