Anti-hallucination validator layer for Claude Code agent teams. Part of a larger 15-agent + 48-skill orchestration system I built as a non-coder over 10 weeks.
Most Claude Code workflows trust the first response. In production, agents hallucinate file paths, import from non-existent packages, and claim completion without verification. That quietly burns hours of debugging.
The validator triad sits after the primary agent(s) and runs parallel checks before the work is trusted:
- skeptic — verifies every claim against real artifacts (files exist, commands output matches, APIs respond)
- reality-checker — validates tasks against the actual codebase (function signatures, imports, exported names)
- completeness-validator — bidirectional traceability between user-spec requirements and shipped code
Plus skill-checker for validating the quality of reusable skills themselves.
primary agents (router / senior-engineer / researcher / ...)
↓
generated output (files, commits, specs)
↓
validator triad runs in parallel
├── skeptic → flags hallucinated references
├── reality-checker → flags spec/code mismatches
└── completeness → flags missing requirements
↓
if any flag → orchestrator retries (max 3) or escalates to user
if all pass → proceed
verification/— production-ready verification protocol (fresh evidence required before "completion" claims)task-decomposition/— atomic task spec generation from approved tech-specsystematic-debugging/— 4-phase root-cause debugging before any fix
These are Claude Code agent + skill definitions. Drop them into your
project's .claude/agents/ and .claude/skills/ respectively. Claude
Code auto-discovers them.
Requires Claude Code (https://code.claude.com).
Extracted from MRZ SWARM — a 15-agent + 48-skill operational framework I built from zero as a non-coder over Feb–Apr 2026, using ~$150 in AI tools. This validator layer is the innovation I rely on most: it's what makes "non-coder ships production" actually true instead of aspirational.
MIT