jarrettmeyer/skills v1.0.4

v1.0.4 of my Claude Code skills plugin is out, with a new gan skill: an adversarial multi-agent loop that iterates until quality converges.

Today I shipped v1.0.4 of jarrettmeyer/skills. If you missed the v1.0.0 post, the plugin is a collection of reusable Claude Code skills — slash commands you can install once and use across all your projects.

claude plugin install jarrettmeyer/skills

The gan skill

The name is borrowed from generative adversarial networks. In a GAN, two neural networks compete: a generator produces samples, a discriminator tries to distinguish them from real data. Neither network sees the other’s internals — just the output. They push each other toward better results through adversarial pressure.

The /jarrettmeyer:gan skill applies the same structure to any artifact, using Claude subagents instead of neural networks.

Two phases

Creation phase. A Creator subagent produces the artifact — a plan, a code change, a memo, whatever you asked for. A Critic subagent then reviews it, with no knowledge of how it was made. The Critic only sees the artifact itself, not the Creator’s reasoning or context. Issues are classified as critical, major, or minor. Critical and major issues trigger another round; minor issues don’t. This repeats until the Critic approves or you hit the round limit.

Execution phase. Once the Creation phase produces an approved artifact, an Executor subagent implements it. A second Critic reviews the implementation against a one-sentence acceptance criterion derived from your original request. Same loop: critical and major issues get another round, minor issues are accepted as-is.

The context isolation is a deliberate design choice. A Critic that can see the Creator’s reasoning will naturally be anchored by it. By passing only the output, the Critic is forced to evaluate the artifact on its own merits.

Invocation

/jarrettmeyer:gan refactor the authentication module

/jarrettmeyer:gan --plan-only draft a build vs. buy strategy to be presented to senior leadership.

/jarrettmeyer:gan --max-rounds 2 --creator-role "novelist" --critic-role "literary editor" write a short story about a recursive function

The --plan-only flag stops after the Creation phase — useful when you want a reviewed plan but aren’t ready to execute. --max-rounds overrides the default of 3 rounds per phase. You can also set custom roles for each agent via --creator-role, --critic-role, and --executor-role.

Not just for code

The skill is general-purpose. Anything with a quality bar can go through it: code, architecture documents, blog post outlines, technical memos, API designs. The role assignment adapts: a coding request gets a software engineer as Creator and a senior code reviewer as Critic; a business case gets a business analyst and a CFO.

The plugin is open source and MIT licensed. All the source is at github.com/jarrettmeyer/skills.