Agent Skills › TNTcraftHIM/Piik › context-consistency

context-consistency

GitHub

用于跨界面、跨层级审查上下文一致性,识别语义差异与规范审计。遵循所有者边界,区分实现与文档差异,确保状态、UI及生命周期的一致性,适用于代码、文案和文档的标准化检查。

.agents/skills/context-consistency/SKILL.md TNTcraftHIM/Piik

Trigger Scenarios

跨层面对比代码与UI/文案一致性 执行仓库级标准化审计 排查重复含义或命名冲突

Install

npx skills add TNTcraftHIM/Piik --skill context-consistency -g -y
More Options

Non-standard path

npx skills add https://github.com/TNTcraftHIM/Piik/tree/main/.agents/skills/context-consistency -g -y

Use without installing

npx skills use TNTcraftHIM/Piik@context-consistency

指定 Agent (Claude Code)

npx skills add TNTcraftHIM/Piik --skill context-consistency -a claude-code -g -y

安装 repo 全部 skill

npx skills add TNTcraftHIM/Piik --all -g -y

预览 repo 内 skill

npx skills add TNTcraftHIM/Piik --list

SKILL.md

Frontmatter
{
    "name": "context-consistency",
    "description": "Review repeated meanings across Piik code, UI, copy, lifecycle and documentation. Use for cross-surface or cross-layer consistency work, recurring fixes in sibling paths, and requested repository-wide standardization audits. Locate common ownership gaps while preserving differences justified by context."
}

Context Consistency

Apply the project's contextual contract. AGENTS.md owns authority and CONTRIBUTING.md owns delivery. This skill is a review method, not a second product specification or permission to change behavior.

Scope And Sources

Read the user's latest accepted decisions and the relevant entries in the owner map. Follow their product, design, engineering or operational owners; do not load every reference for a scoped fix. Treat desired rules, current implementation and verified behavior separately. An undocumented preference or an old audit is not an accepted contract.

For a scoped change, follow the affected semantic family and its actual callers. For a requested repository-wide audit, first inventory standards and reachable surfaces across Browser/App/Server, shared interfaces, website/demos, tooling and guides. Record examined and unexamined areas so a large finding count cannot stand in for coverage. Review-only work must not alter product sources.

Trace Before Comparing

Follow a concrete producer through state/normalization, effects, consumers and cleanup. Record the subject, asserted fact or requested outcome, authority, scope/generation, lifetime and user intent. Use a small comparison table only when there are multiple variants to compare.

Ask whether an apparent inconsistency is:

  • one meaning implemented or documented differently;
  • different meanings accidentally using one label, flag or representation;
  • a missing, conflicting or bypassed shared rule;
  • an intentional adaptation supported by capability, authority or user context.

Compare timing as well as steady appearance: initial state, update, replacement, failure/recovery and retirement. For async work, check A replaced by B before A's completion or cleanup. For presentation, check actual state against copy, icon, comic, mascot, motion and announcement, including simultaneous surfaces. Do not infer progress from decoration, availability from configured capability, or media readiness from signaling alone.

Correct The Owning Boundary

When fixes are authorized, choose the smallest coherent correction at the existing owner and update its reachable consumers. Share implementation where the semantics and lifetime match; use narrow adapters for real runtime or presentation differences. A similar name or repeated syntax is not sufficient reason to merge states or introduce a generic manager.

If the intended semantics are unresolved, put that decision in TODO and continue independent checks. Do not turn a proposal into a mandatory rule merely by adding it to a table.

Verify the changed producer-to-consumer boundary with existing previews, tests or bounded runtime checks. Check that the correction preserves justified differences and removes obsolete writers, mappings or timers. Stop broadening verification after relevant checks pass unless new evidence warrants it.

Deliver Useful Findings

Group occurrences by their root ownership or contract issue. Give each confirmed finding its intended rule, reachable trigger, evidence, affected scope, smallest correction and regression risk. Label unverified suspicions and design choices separately. Do not claim a repository-wide pass from static searches or one view.

Use the existing owner map as the index of standards. Consolidating files must preserve ownership, inbound links, tool-discovered paths and contributor entry points; test those references after moves. Keep architectural rationale, user guides and temporary audit evidence distinct from normative rules.

Absorb accepted lessons into their existing owner, keep remaining work in TODO, and leave completed history in Git/PRs. Raw inventories, screenshots and audit reports remain ignored working artifacts. Do not add a parallel permanent audit ledger, scoring system, new dependency or extra approval flow.

Version History

  • a9b7507 Current 2026-09-21 23:49

Same Skill Collection

.agents/skills/ponytail/SKILL.md

Metadata

Files
0
Version
04cbb2c
Hash
4af29b9a
Indexed
2026-09-21 23:49

Главная - Вики-сайт
Copyright © 2011-2026 iteam. Current version is 2.155.2. UTC+08:00, 2026-10-03 02:39
浙ICP备14020137号-1