Show HN: Claude Forge – GAN Inspired Adversarial Pipeline

github.com · hatmanstack · 1 day ago · view on HN · tool
quality 2/10 · low quality
0 net
AI Summary

Claude Forge is an adversarial pipeline using five Claude Code agents in generator/discriminator loops (Planner, Plan Reviewer, Implementer, Code Reviewer, Final Reviewer) to improve code correctness through iterative feedback and review cycles, trading 3x tokens and time for higher quality outputs.

Tags
ai-assisted-development code-generation llm-engineering adversarial-pipeline claude-api code-review-automation gan-inspired prompt-engineering test-driven-development multi-agent-systems
Entities
Claude Claude Forge Claude Code
Five Claude Code agents in generator/discriminator loops: Planner<->Plan Reviewer, Implementer<->Code Reviewer, Final Reviewer. Each agent gets a fresh context window, discriminators only see output, never reasoning.

Plans are immutable; feedback routes through a separate file. Reviewers probe for hallucinated paths, dependency cycles, ambiguous specs, false-positive test strategies. Implementer does TDD with atomic commits. Max 3 iterations per loop before human escalation. Everything persists to disk, resume mid-pipeline if you hit token limits.

Two skills: /brainstorm (interactive spec generation grounded in codebase exploration) and /pipeline (automated build cycle). Drop .claude/skills/ into any project.

Trade-off: ~3x tokens and wall time vs. single-pass. Worth it when correctness matters.

← Back to stories