Gan Trinity — AI skill for Claude Code

GAN Trinity · 三体

[English](#english) · [中文](#chinese)

---

English

**Three AI agents that verify each other — so the one writing the code is never the one grading it.**

A multi-agent development workflow skill for **WorkBuddy** and **Claude Code**. It splits a complex task into sprints, and every sprint is handled by three isolated agents in sequence:

Planner (explores the repo) → Contract negotiation → Generator (implements)
    → Evaluator (independently verifies) → Fix loop

Each role runs as a **separate subagent with its own context**. The Evaluator cannot see the Generator's reasoning, so it has no way to rubber-stamp its own work. That's a structural guarantee, not a prompt-level request.

Install

# WorkBuddy
cp -r gan-trinity ~/.workbuddy/skills/

# Claude Code
cp -r gan-trinity ~/.claude/skills/

Use

/gan-trinity 

Or just say *"use the trinity workflow to build X"*.

The part that actually matters

Before any code is written, the **Generator proposes acceptance criteria and the Evaluator edits, extends, or rejects them**. This kills ambiguity before implementation starts — and it catches a class of errors that self-review structurally cannot.

Real example: during contract review, the Evaluator blocked a data bug where a generator had mistaken a **filename suffix for a file size** (`Photoshop Temp151073331068` — that trailing number is part of the filename, not the byte count). The wrong derivation happened to produce a plausible-looking result, so no amount of self-checking would have found it. See the [full case study](examples/disk-cleanup-case.md).

When to use it

Good for: scripts, configs, documents, batch jobs, design artifacts — anything that deserves independent quality control.

Not for: typo fixes or one-line changes. Each sprint spawns at least 4 subagents, so it's expensive by design.

**Verification status (tested 2026-09-21)**: Working on WorkBuddy ac