// AI-assisted development of anti-hallucination specifications. Check out the seven-layer defense architecture, quantitative indicators, execution templates and practical cases. Used for code review, task startup inspection, and weekly review.
Audit AGENTS.md / CLAUDE.md against the five high-leverage patterns (progressive disclosure, procedural workflows, decision tables, production code examples, domain rules with concrete alternatives). Reports per-pattern coverage, anti-patterns, and a prioritized fix list.
Shell 脚本中 awk 的 POSIX 兼容性指南。 Use when: 编写或审查包含 awk 的 shell 脚本， 尤其是需要 macOS + Linux 跨平台运行的场景。 触发词: awk, BSD awk, POSIX regex, [[:space:]], guard 脚本, 跨平台 shell
Assessment-driven development — Quantify code generation quality with pass@k / pass^k metrics, automatically scored by Grader.
Iterative retrieval — 4-stage loop (DISPATCH→EVALUATE→REFINE→LOOP) to pinpoint relevant information in the code base. Up to 3 rounds.
Strategic compression — Manual compression of contexts at logical boundaries rather than arbitrary automatic compression. Key decisions and constraints are preserved and intermediate exploration is discarded.
Post-hoc diagnosis of a failed agent trajectory. Classifies the first unrecoverable step into one of nine failure categories (plan adherence, hallucinated information, invalid tool call, misread tool output, intent–plan mismatch, under-specified intent, unsupported intent, guardrail trigger, system failure) and produces an evidence-backed root-cause report.
| name | vibeguard |
| description | AI-assisted development of anti-hallucination specifications. Check out the seven-layer defense architecture, quantitative indicators, execution templates and practical cases. Used for code review, task startup inspection, and weekly review. |
#VibeGuard — Anti-hallucination specification Skill
VibeGuard is an anti-hallucination framework for AI-assisted development that systematically blocks common failure modes in LLM code generation through a seven-layer defense architecture.
Canonical contract sources for this skill:
README.md — product entry and current Core vs Workflow boundarydocs/rule-reference.md — public rule/guard summaryschemas/install-modules.json — install/runtime contractdocs/internal/history/spec.md remains a historical design snapshot and should not be treated as the authoritative implementation contract.
Calling /vibeguard can:
Triggered when user mentions:
| Hierarchy | Name | Key Tools/Rules |
|---|---|---|
| L1 | Anti-duplicate system | check_duplicates.py / Search first then write |
| L2 | Naming constraints | check_naming_convention.py/snake_case |
| L3 | Pre-commit Hooks | ruff / gitleaks / shellcheck |
| L4 | Architecture guard testing | test_code_quality_guards.py Five rules |
| L5 | Skill/Workflow | plan-flow / fixflow / optflow |
| L6 | Prompt embedded rules | CLAUDE.md mandatory rules |
| L7 | Weekly review | review-template.md |
Refer to references/task-contract.yaml and confirm:
1. Goals are clear and verifiable
2. The data source has been determined
3. Acceptance criteria can be tested
Refer to references/scoring-matrix.md to score each finding:
- impact: 1-5
- effort: 1-5
- risk: 1-5
- confidence: 1-5
Formula: priority = (impact × confidence) - (effort + risk)
Refer to references/review-template.md, record:
1. Return event this week
2. Guard interception statistics
3. Indicator trends
4. Highlights for next week
references/task-contract.yaml — Task startup Checklist (machine verification format)references/review-template.md — weekly review templatereferences/scoring-matrix.md — risk-impact scoring matrixREADME.md (repository root directory) — current product entrypointdocs/rule-reference.md (repository root directory) — current rule/guard summaryschemas/install-modules.json (repository root directory) — current install/runtime contract