From inside Claude Code — /ultraexpertise

Run oe ultra without leaving your Claude Code session — install once, invoke anywhere.

When you need this

• You are already in a Claude Code conversation and want to author an experience without switching to a terminal.
• You want Claude Code to walk you through open questions interactively after the draft lands.
• You want schema validation errors explained in plain English rather than raw AJV output.
• You are exploring whether a given workflow is a good fit for OpenExpertise before committing.

The minimal example

Install the slash command once:

mkdir -p ~/.claude/commands
cp packages/skill-experience-creator/commands/ultraexpertise.md ~/.claude/commands/

Then in any Claude Code session:

/ultraexpertise Review my repo for OWASP Top 10 issues

Claude Code runs oe ultra with the task text, reports back the slug and draft path, lists open questions, and tells you whether validation passed.

How it works

The file at packages/skill-experience-creator/commands/ultraexpertise.md is a Claude Code slash command definition. When invoked, Claude Code reads the markdown, substitutes $ARGUMENTS with whatever you typed after /ultraexpertise, and executes the instructions as a system prompt.

The command instructs Claude to:

1. Check that oe is on PATH. If not, it suggests building from the workspace: pnpm -r build && node packages/cli/dist/bin.js ultra "...".
2. Pick the draft root. If an experience.yaml is nearby the slash command defaults to .openexpertise/drafts/; otherwise it uses ./openexpertise-drafts/.
3. Run oe ultra "<task>" --draft-root <dir>. The full output — slug, phases, node kinds, open questions, validation result — is surfaced back to you in the chat.
4. Summarize next steps. oe run <draft path>, the mv to promote, and any open_questions[] items that need answering.
5. Explain validation failures. If validation fails, Claude reads the generated experience.yaml and describes the schema error in plain English, with an offer to fix it.

Claude will not run oe ultra if neither ANTHROPIC_API_KEY nor OPENAI_API_KEY is set — it checks first and asks you to set the right key.

Variations

No oe binary yet — use the workspace path:

If the CLI hasn't been installed globally, Claude Code falls back to:

node $(pwd)/packages/cli/dist/bin.js ultra "Your task here"

Force a provider:

Type a flag inline:

/ultraexpertise --llm openai Classify support tickets by priority

Claude Code passes it through to oe ultra.

After the draft lands, ask follow-up questions in the same session:

The draft has a fetch_pr node — what env var should I set for the GitHub token?

Claude knows the draft structure and can advise without re-running the command.

Gotchas

• The command file must be in ~/.claude/commands/, not ~/.claude/ or a project-local directory, for Claude Code to discover it as a slash command.
• Claude Code will not auto-promote the draft. Promoting (mv draft examples/slug) is always a user action — the command explicitly forbids it.
• Editing the generated draft in the same session is advisory only. Claude can point out problems and suggest fixes; applying them requires you to edit the file manually or run oe ultra again.
• $ARGUMENTS is the raw text after /ultraexpertise. Quotes are not required but avoid unescaped single quotes in your task description.

See also

• oe ultra — LLM authors for you — the underlying CLI command
• MCP server — the oe_ultra MCP tool (for Codex, Gemini, and other MCP clients)
• Skills + SKILL.md — the SKILL.md format the slash command is built on