agent-skills

This is a public repository. Do not commit API keys, passwords, tokens, internal URLs, or any credentials. Use .env for secrets and verify .gitignore is working before every commit.

Portable agent skills for delegating work to subagents, verifying what they return, and authoring skills of your own. Each skill is a folder with a SKILL.md and its reference files. They load unmodified in any harness that scans a skills directory; the name + description frontmatter is the only contract a consumer reads.

What's here

Delegation and verification:

• delegation-sizing: decide whether a task is worth handing to subagents, which pattern fits, and how much wall-clock a round saves.
• briefing-subagents: write a brief that makes returned work auditable.
• worktree-isolation: spawn parallel file-writing subagents without collisions.
• verifying-delegated-output: check claims, git state, and content before you trust returned work.
• dynamic-workflows: author a Claude Code dynamic workflow, or decide when a workflow beats manual delegation.

Authoring:

• meta-skill-creator: the doctrine and builder for writing portable single-job skills.

Install

Subscribe through the Claude Code plugin marketplace, so merged updates reach you on the next session:

/plugin marketplace add Exploitacious/agent-skills
/plugin install agent-skills@agent-skills

Or copy the files into any harness. Clone the repo and drop any skills/<name>/ folder into the directory your harness scans for skills. You own the copy from then on, updates included.

License

MIT. See LICENSE.

Credit

The skill shape and progressive-disclosure model follow the agentskills.io spec. The prose follows the unslop writing style vendored from cursor/plugins pstack/skills, whose model in turn draws on mattpocock/skills writing-for-agents.