Agent skill

The Iris agent skill is a SKILL.md with references for writing, explaining and debugging reader specs. It covers Iris 0.1. The skill has the agent validate its work with the CLI.

Install from the repository

The source is skill/iris/. Copy that whole directory, including SKILL.md and references/, into your agent’s skills directory. For example, a project-local Claude Code installation places it at .claude/skills/iris/; a project-local Codex installation places it at .agents/skills/iris/. Keep the reference paths relative to SKILL.md intact. Build the CLI and make iris available on the agent’s PATH.

A packaged Claude Code plugin is planned; copying the directory is the installation path for now.

What it covers

The workflow starts with the format description and sample input, looks for a close match in the format library, then drafts a reader spec. The agent checks the spec, reads the sample, compares every entry with the input, adds ranges and checks for the model’s assumptions, tests invalid and edge-case inputs, and formats the result.

The references cover reading patterns, exact language rules, diagnostics, existing formats, and embedding. They address common mistakes: counts that mean records rather than values, line-aware statements that advance unexpectedly, sentinels, array columns from repeats, checks over records, and instance selection. They also describe what Iris 0.1 cannot express and when the host program must finish validation.

The Specification is normative. When skill guidance and the Specification disagree, follow the Specification. See the error reference for stable language error codes.