Skills

Going Dark: A Strict Method for Getting AI Coding Agents from Issue to Release

If you want AI coding agents to run your software project through one strict pipeline — issues written as specs, handoff prompts for any agent, audits that fix what they find — Going Dark is for you. First confirm whether your agent is Claude Code (marketplace path) or another agent that reads AGENTS.md (npx installer); the two install paths are completely different.

Official source · going-dark

Install

/plugin marketplace add reyxdz/going-dark
/plugin install go-dark@going-dark

Who It's For

  • For people who maintain software projects with AI coding agents (Claude Code, OpenAI Codex, Gemini CLI, GitHub Copilot, Cursor, and any agent that reads AGENTS.md) and want to treat issues as specs, drive any agent with handoff prompts, and get audits to actually fix what they find.
  • This is "a strict way of doing things": one unified pipeline from issue to release, not a set of unrelated tools. If you only want a plugin for one specific feature, or your project doesn't work in an issue/spec-driven way, it may not be a good fit.
  • The official evidence does not state any additional prerequisites regarding project language, operating system, or team; these are unknowns you should verify yourself.

Your First Attempt in 3 Minutes

The recommended main path is installing the plugin in Claude Code:

  1. Prerequisite: Claude Code is installed. Node.js 20+ is only required on the npx install path (the evidence for the Claude Code marketplace path makes no mention of a Node requirement).
  2. Run the following in Claude Code:
/plugin marketplace add reyxdz/going-dark
/plugin install go-dark@going-dark
  1. Trigger a low-risk task: after installation, standing rules take effect in the session, such as "minimal local load — read the code, rely on CI, run only the narrowest checks locally, and stop anything you started," plus never force-push and never print/copy/commit/screenshot secrets. You can first ask the agent to read the code and produce an issue-style spec or audit findings without letting it modify files.
  2. Success signal: the standing rules above (such as minimal local load, git safety, and secret protection) appear among the agent's session rules, and it responds to your requests in an "issue as spec" manner.

In other agents, each skill is installed as go-dark-<skill>, with rules written into that agent's rules file as a marked block; the installer refuses to overwrite folders it did not create, and uninstall only removes what it added. Beyond that, the exact skill names, operation lists, and per-agent parameters are not provided in the official evidence; check the repository README to confirm.

Use Cases

Case 1: Turn a bug into a spec. You find a bug in your project with clear reproduction steps. Tell the agent: "Using the go-dark method, write this bug up as a spec-style issue. Don't fix anything." It follows the issue-as-spec flow and outputs a spec document that any agent can execute. The confirmable outcome is a structured issue text; whether an issue is then filed automatically, and into which file, is not stated in the evidence — confirm with the agent before execution.

Case 2: A diagnosis-only audit. Ask the agent: "Audit this module, report findings only, don't fix anything yet." It follows the "audits fix what they find" flow, but your instruction makes it diagnosis-first. The expected output is a list of audit findings; because the standing rules require minimal local load and only the narrowest checks, the agent will most likely focus on reading code and rely on CI. Which specific check commands will run, and whether it accesses the network, is not stated officially and is an unknown.

Boundaries

  • Distinguish the install paths clearly: Claude Code uses the two marketplace commands; other agents use the npx installer (requires Node.js 20+, no global install), with skills named go-dark-<skill> written as marked blocks into the rules file. The marketplace install method cannot be applied to other agents.
  • Side effects and confirmation points: outside Claude Code, the installer creates folders and modifies the agent's rules file; it will not overwrite folders it did not create, and uninstalling only removes what it added. Beyond that, which files the skills read at runtime, which commands they execute, and whether they access the network are not stated in the official evidence — before use, have the agent first explain what it is about to do.
  • Secret protection is an official standing rule: never print, copy, commit, or screenshot secrets; never force-push to git. These are method-level constraints, not hard filesystem isolation.
  • Not yet confirmed: the exact skill list and names, per-agent install parameters, supported project languages, and the detailed steps of the audit and handoff flows. Before use, read the repository README directly to verify.