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.