Agent Skills
› mtarcure/claude-vibe-squad
› layered-analysis-loop
layered-analysis-loop
GitHub一种深度代码分析技能，通过结构、行为、边界和矛盾四个分层迭代进行系统性审查。每层聚焦特定问题并记录证据驱动的变更，直到模型收敛或达到预算限制，旨在发现深层缺陷而非浅层总结。
Trigger Scenarios
需要对复杂代码或系统进行深度结构化分析
常规阅读无法覆盖边缘情况或潜在矛盾
Install
npx skills add mtarcure/claude-vibe-squad --skill layered-analysis-loop -g -y
SKILL.md
Frontmatter
{
"name": "layered-analysis-loop",
"audience": "specialist",
"description": "Use when analyzing a target deeply enough to require distinct structural, behavioral, edge-case, and contradiction passes rather than one undirected read—fix a question for each layer, record its evidence-backed delta, and stop only at a no-delta layer or an explicit budget cutoff."
}
Layered Analysis Loop
Analyze in deliberate passes that each answer one question, so depth accumulates on a stable base instead of one undirected read producing a shallow summary.
Steps
- Write the question each layer will answer before starting it. A layer without a question becomes a re-read.
- Run layer 1 for structure only: what exists, how it is organized, where the entry points are. Resist diagnosing anything yet.
- Run layer 2 for behavior: follow the primary paths end to end and record what actually happens, in order.
- Run layer 3 for edges: error paths, empty and boundary inputs, concurrency, and the states the primary path assumes but does not enforce.
- Run layer 4 for contradiction: actively try to falsify the model built by layers 1-3. Look for the case that breaks it rather than the case that confirms it.
- Close each layer with a written delta — what changed in the model — and carry forward only findings with evidence. An unrecorded layer did not happen.
- Stop when a full layer produces no delta, and say so. Continuing past convergence spends budget; stopping before it ships an untested model.
- Report the layers run, the delta from each, and the layer at which convergence occurred.
Acceptance
- Each layer has a written question fixed before the layer ran.
- Structure, behavior, edge, and contradiction layers are separately recorded.
- Every layer closes with an explicit delta, including "no change".
- The contradiction layer names at least one specific attempt to falsify the model.
- The stopping condition is stated as convergence or as an explicit budget cutoff.
Version History
- d5262e2 Current 2026-09-11 11:25
