Usage

Invoking the skill

Claude Code — type the slash command when starting Olympus work:

/shipd-olympus

Claude also auto-loads it when your prompt mentions shipd, Olympus, or submission-shaped work (the skill's description field triggers on those).

Cursor — the rule is agent-requested: mention "shipd olympus submission" in your prompt, or @-mention the rule.

AGENTS.md agents — the pointer added to AGENTS.md tells the agent to read skills/shipd-olympus.md when working on a submission.

What to ask for

The skill turns these prompts into rule-compliant work instead of naive attempts:

Find me an eligible repo for a shipd Olympus submission

→ agent checks license (by reading LICENSE, not the GitHub API field), stars, activity window, language, and warns about quest-farm repos.

Design a task on <repo> that lands under the 50% pass bar

→ agent picks a complex subsystem and stacks difficulty archetypes (convention traps, exact-boundary rules, deliberate sibling differences, wide wiring) instead of hoping the agent batch comes back hard.

Sweep this idea for originality: <idea> on <repo>

→ agent searches PRs in all states (closed PRs kill ideas too), issues, and GitHub Discussions.

Write the test patch for this task

→ hash-suffixed filenames, typeof-guards on throws tests, no error-message pinning, broad base-mode ignore glob.

Verify my submission like a shipd reviewer would

→ fresh clone, docker build, --network none run, base-pass/new-fail/apply/both-pass loop, leak-word scan.

When checks fail

Paste the raw check output into the chat. The skill contains the known checker behaviors: which Dockerfile warnings are non-blocking and why, how to respond to description request_changes rounds without trimming contract-critical facts, and the test-name collision fix.