Skills

ship-ready: Ship AI-written Code with Evidence

ship-ready suits developers using Claude Code, Codex, Copilot, or Gemini CLI who want AI-written code turned into an evidence-backed "ready to ship" state—provided you first run the official install script and accept it writing a guardrails marked block into the agent instruction files.

Official source · ship-ready

Install

git clone https://github.com/notaaaronfr/ship-ready.git && cd ship-ready && ./install.sh --all

ship-ready is an Agent Skill that makes AI-generated code safe to ship: it locks down behavior, finds bugs and security vulnerabilities, fixes Big-O, adds property and mutation tests, and refuses to declare "done" until evidence is in hand. This guide helps you decide whether it fits your workflow and run your first try.

Who it's for

  • Developers who write code with Claude Code, OpenAI Codex, GitHub Copilot (CLI and VS Code), or Gemini CLI / Antigravity, and want an evidence-backed "ready to ship" check after those agents finish writing code.
  • Official documentation also states it works with any agent that reads AGENTS.md, and can be pasted in as a single file.
  • Clear environment prerequisites: macOS or Linux with bash and git installed; on Windows, use WSL or Git Bash.
  • When it's not a fit: official materials do not state which programming languages, project types, or codebase sizes it supports—these are unknowns and cannot be guaranteed on your behalf. Also avoid trying it in environments without bash/git.

Your first run in 3 minutes

Prerequisites: macOS or Linux (Windows via WSL or Git Bash), with bash and git installed; at least one officially supported agent on your machine (e.g., Claude Code).

The officially recommended main path is installing to all supported agents via the repository's install script:

git clone https://github.com/notaaaronfr/ship-ready.git && cd ship-ready
./install.sh --all

If you only want it for specific agents, the official docs also offer ./install.sh --target claude,codex; this is an officially supported optional variant, not a second main path alongside the first.

Installation produces two outcomes: first, skill files are placed per agent—Claude Code gets ~/.claude/skills/ (user-level installs are symbolic links), while Codex/Copilot/Gemini share ~/.agents/skills/; second, the install script writes a marked guardrails block into the corresponding agent's instruction file (CLAUDE.md, AGENTS.md, GEMINI.md, or ~/.copilot/copilot-instructions.md). This is a file-modifying side effect: it writes to those files but only touches its own marked block, leaving the rest of the file untouched. If you don't want this persistent guardrail, opt out at install time with --no-guardrails.

Once installed, give your agent a low-risk check task in an existing project—for example, ask it to run a diagnostic evaluation with ship-ready on code an AI just wrote. Per the official description, its philosophy is "no evidence, no done," so the success signal is concrete output: behavior lock records, bug/security findings, or test evidence—rather than a vague "all done."

To be clear: official evidence does not detail, check by check, which files it reads, whether it executes commands, or whether it accesses the network—these behaviors are unknown. Before letting it run, it's wise to understand what it will do to your repo and confirm your code is committed or backed up.

Use cases

Case 1: Review a function an AI just wrote. While developing with Claude Code or Codex, the agent generated some logic; you then issue the task: "Use ship-ready to check this function—lock behavior, find bugs and security issues, and add tests." Per the official capability description, you can expect behavior lock records, bug/security vulnerability findings, and evidence in the form of property and mutation tests—not an unverified "done" claim.

Case 2: Verify before and after performance optimization. You suspect a piece of code has a complexity problem; have the agent fix Big-O with ship-ready while keeping behavior locked before and after the change. The official docs explicitly list "fixes Big-O" as part of it, so you can be confident it addresses complexity; whether you need to manually run benchmark scripts yourself is not stated in the evidence and remains unknown—it's best to spell out the verification method in your task description.

Limits and boundaries

  • Installation itself modifies files: besides placing skill files, the script writes marked blocks into CLAUDE.md / AGENTS.md / GEMINI.md / copilot-instructions.md; user-level installs are symbolic links, which official docs note are updated by git pull. Confirm this before running the install script.
  • Install scope is determined by flags like --all or --target claude,codex; only the agents you install for get the skill. Don't assume every tool on your machine will be affected.
  • Language and platform boundaries: official docs do not state which programming languages are supported, the scope of text/code languages, or any additional limits beyond the runtime requirements—none of this is confirmed.
  • The exact runtime behavior around file reads/writes, command execution, and network access is not described in official evidence and is unknown. Before pointing the skill at production code, a trial run in a recoverable environment is the safer approach.