Responsible Agent Skill Development

Responsible-Agent-Skill-Development is a draft (version 0.1) voluntary code of practice for authors of Claude and agentic-AI skills, instruction-based units that direct a general-purpose model to carry out tasks, often with browser or HTTP tool access. The project addresses a dual-use problem: a skill that automates interaction with a third-party website typically contains both runtime information, such as code, selectors, and request shapes required to execute, and research information, such as observed anti-bot checks, rate-limit thresholds, and challenge trees. The code proposes keeping these in separate workspaces, with a private development workspace holding unfiltered research artefacts and a public published workspace containing only runtime-necessary material. It also specifies five documentation filters, including auth-narration, bypass-framing, dev-comment, naming, and fixture filters, along with a nine-item pre-publication checklist designed for an AI agent to run against a diff.

The repository is written in Typst and released under the MIT licence. It contains a banner asset, the code text, and paste-ready markdown snippets that skill authors can add to their README or CONTRIBUTING files to signal adoption, alongside a suggested machine-readable marker file named .responsible-skill-development.yaml for tooling. The document states explicitly what it does not cover, namely certification, disclosure of outright security flaws, and legal compliance, and it positions itself as an addition to ordinary engineering practice rather than a replacement for it.

The project is likely to interest developers and maintainers who publish browser-automation or web-interacting agent skills, as well as those reviewing such skills for publication. It may also be relevant to researchers studying norms for the release of agent tooling. Because the code is self-attested and non-binding, its practical effect depends on voluntary adoption by skill authors.