ship-check coordinates the specialist ai-shipping skills in order: document the system, wire the agent’s operating instructions, review correctness, run security and performance audits in parallel, hand the change to a fresh independent reviewer, derive the test-coverage map, and compile a shipping packet.
It starts only when the user types it, because it writes docs, edits agent instructions and runs a long session. Before writing anything it checks that the checkout is current and asks before overwriting records the owner maintains.
What You Get
- System docs created or refreshed through document-app, and agent instructions (CLAUDE.md or AGENTS.md) derived from them
- Correctness, security and performance findings summarized, with the model mix each audit ran on
- An independent review round from a fresh read-only session, with every finding verified by hand against the code
- A test-coverage map and one shipping packet that lists launch blockers and recommended next actions
Common Use Cases
- A first release of an AI-built product that needs a full pass before launch
- A release review where a human signs off on one packet
- Checking a big refactor against documented intent
- Preparing an app for outside reviewers or customers’ security questions
What Makes It Unique
It sequences existing specialist skills and synthesizes their results, so each stage can also be run alone. The independent review step is mandatory once the full pass runs, and the packet records which reviewer ran and how independent it was. A fetch that shows missing commits stops the pass, and existing owner-maintained docs or coverage maps trigger a question before any overwrite.
Under The Hood
The skill is markdown instructions plus a copy of the independent-review wrapper script, review.py, and an agents folder. The agent runs git fetch and log commands, reads the repository, writes docs, agent instruction files and report files, and starts parallel subagents. For the review it runs review.py, which launches a fresh read-only Codex, Claude Code or Grok session and writes the subject and prompt files into the git directory. It never pulls or merges. It treats the reviewed repository as untrusted input.