AI Agent Skills

Iterion ships two Agent Skills, compatible with Claude Code, Codex, Cursor, Windsurf, GitHub Copilot, Cline, Aider, and other AI coding agents. Install the one you need — they are independent.

Skill	Teaches your agent to
iterion-dsl	Author, review and debug .bot workflows in the current DSL
adversarial-review-loop	Break its own change before a reviewer does — repo-agnostic, no iterion required

iterion-dsl — author workflows

npx skills add https://github.com/SocialGouv/iterion --skill iterion-dsl

The skill teaches an agent to author workflows; its operating counterpart is the MCP server (iterion mcp), which gives the same agent typed tools to run Iterion — launch and follow runs, answer questions, groom the board, and drive a cloud instance. Install both for the full loop: write a bot, then operate it without leaving the agent.

File	Content
SKILL.md	Complete DSL reference — node types, properties, edge syntax, templates, budget, MCP
SKILL-run-and-refine.md	Practice guide for running, debugging and iteratively refining .bot workflows against real data
references/dsl-grammar.md	Formal grammar specification (EBNF)
references/patterns.md	10 reusable workflow patterns with annotated snippets
references/diagnostics.md	Authoritative sparse catalogue of DSL diagnostics (C001–C199 plus async C240–C242) and bundle checks (C200–C234), with causes and fixes

Once installed, just ask your agent to write workflows:

• "Write a .bot workflow that reviews a PR with two parallel reviewers"
• "Create an iterion pipeline that fixes CI failures in a loop"
• "Add a human approval gate before the deployment step"

The agent will use the skill reference to produce valid .bot files that pass iterion validate.

adversarial-review-loop — ship a change through a loop

npx skills add https://github.com/SocialGouv/iterion --skill adversarial-review-loop --full-depth

--full-depth is required here, and it is not optional politeness: the skills CLI stops at a repository's root SKILL.md when it finds one, and iterion has one (that is iterion-dsl). Without the flag the CLI reports "Found 1 skill" and this one is invisible. Verify either way with --list.

A single file — skills/adversarial-review-loop/SKILL.md — and nothing in it is iterion-specific: it works in any repository, with any review gate or none.

It is the practice this project ships its own changes through: a subagent whose posture is to refute the diff rather than bless it, a verification pass that judges the subagent as harshly as it judged the code, fixes applied at the class rather than the site, tests proved by mutation, and three named exits for a loop that has stopped converging. It carries a round budget (a ceiling on cost, never a target), the commit trailers that report what a review cost, the upstream plan review by a model of another family, and the journal/retrospective discipline that keeps the protocol honest instead of dogmatic.

How iterion itself applies it — where the round is required, the ceiling by change size, who pays a local round versus a gate cycle — is in docs/agents/workflow/adversarial-review-loop.md.

Installing from iterion itself

iterion skill import installs a public skill pack as an enable/disable-able plugin, with no npx involved:

iterion skill import https://github.com/SocialGouv/iterion
iterion plugin enable iterion               # packs install disabled by default
iterion --json plugin info iterion          # lists what it contributes

It collects the markdown under the repository's skills/ directory, so this installs adversarial-review-loop. The iterion-dsl skill lives at the repository root, next to its references/, and travels through npx skills add above.

Do not check the result with iterion skill list: that command reads the skill library (iterion skill add), a different store, and it answers "No skills in the library" for a perfectly installed pack. A plugin's skills show up under iterion plugin info, and in the workspace's .claude/skills/ once a run starts.

The same command takes any third-party skill repository (iterion skill import https://github.com/acme/awesome-claude-skills) or a local directory holding a skills/ tree. Once enabled, a plugin's skills are mirrored into the workspace's .claude/skills/ at run start in both discovery shapes — the directory form <name>/SKILL.md (what claude_code's Skill tool discovers) AND the flat alias <name>.md (what a prompt Reading the skill by path resolves; claw reads both). The same dual-shape mirror applies to the skill library (iterion skill add), so all three tiers — bundle, plugin, library — land the same shape and marker layout. See docs/plugins.md and ADR-079.

For claude_code's own Skill tool, which discovers only the <name>/SKILL.md directory form, prefer npx skills add above: it installs the directory form directly.