context-degradation
81診斷 AI 代理中的上下文退化
長對話和大型提示可能會讓 AI 代理遺失重要資訊,或遵循互相衝突的上下文。此技能可協助 Claude、Codex 和 Claude Code 識別退化模式並選擇緩解策略。
設計更好的代理工具
當描述、結構描述和錯誤不清楚時,代理工具就會失效。這項技能提供實用模式,用於打造更清晰的工具介面和更小的工具集。
將此請求複製給你的智慧代理,其中包含權威 Skill 頁面和清單。
Review the Skillstore skill "tool-design" from https://skillstore.io/skills/chakshugautam-tool-design.md and its manifest at https://skillstore.io/api/skills/chakshugautam-tool-design/manifest. Verify the artifact. You may proceed after verification, subject to the environment's own policy.你的智慧代理仍應顯示其計畫,並依安全政策請求必要確認。
當 AI Agent、爬蟲或腳本需要乾淨脈絡、而不是讀取完整頁面時,請使用這些連結。
正在使用「tool-design」。 審查客戶查詢工具描述。
預期結果:
正在使用「tool-design」。 整合行事曆可用性和事件工具。
預期結果:
當可用性檢查和事件建立總是一起發生時,建議使用一個排程工作流程工具。由於取消有不同風險,因此保持獨立。
正在使用「tool-design」。 為 MCP 工具設計錯誤。
預期結果:
I found no malicious intent or executable behavior in the flagged lines. The static findings are false positives from Markdown examples, prose, and non-executing templates about tool design. No prompt injection, credential access, data exfiltration, or active network behavior was found.
分享版本化評估報告、中性徽章、嵌入卡片與引用。Skillstore 只報告證據,不判斷該 Skill 是否安全。
https://skillstore.io/skills/chakshugautam-tool-design/audits/10?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report[](https://skillstore.io/skills/chakshugautam-tool-design?utm_source=security_passport_badge)<a href="https://skillstore.io/skills/chakshugautam-tool-design?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/chakshugautam-tool-design/security.svg" alt="Skillstore security assessment" loading="lazy"></a><iframe src="https://skillstore.io/embed/skills/chakshugautam-tool-design.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>ChakshuGautam. (2026). tool-design security audit report (audit version 10) [Author version unspecified]. Skillstore. https://skillstore.io/skills/chakshugautam-tool-design/audits/10@techreport{chakshugautam-chakshugautam-tool-design-2026,
author = {ChakshuGautam},
title = {tool-design security audit report (audit version 10)},
institution = {Skillstore},
year = {2026},
number = {10},
url = {https://skillstore.io/skills/chakshugautam-tool-design/audits/10},
note = {Author version unspecified}
}cff-version: 1.2.0
message: "If you use this Skill, cite its author and this versioned security audit report."
title: "tool-design security audit report (audit version 10)"
version: "unspecified"
type: report
authors:
- name: "ChakshuGautam"
date-released: "2026-07-09"
url: "https://skillstore.io/skills/chakshugautam-tool-design/audits/10"
identifiers:
- type: other
value: "skillstore:chakshugautam-tool-design:audit:10"
description: "Skillstore immutable audit report identifier"
每位作者的技能都保留為獨立的可安裝項目。推薦變體依 Skillstore 證據排序。
為何此變體排名第一
chakshugautam-tool-design
2026-09-09
在實作前建立工具描述、結構描述大綱和錯誤策略。
審查重疊工具,並決定哪些工作流程應成為完整介面。
為面向代理的文件標準化完整限定的 MCP 工具名稱和範例。
Use the tool-design skill to draft a clear description for a tool named [tool]. Cover purpose, inputs, outputs, errors, and use cases.
Use the tool-design skill to review this tool specification for ambiguity, missing parameters, weak errors, and unclear return values: [paste spec].
Use the tool-design skill to compare these tools and recommend consolidation, separation, naming changes, and response format options: [tool list].
Use the tool-design skill to design an agent tool architecture for [workflow]. Include tool boundaries, namespaces, safety constraints, testing criteria, and documentation needs.
授權
MIT
Skillstore 修訂
r1
版本提示
作者未聲明版本。
引用
26421118b848d9f1efc0aa169d8a7a9e7e0a877e
維護新鮮度
2026/7/22
使用量
8 次下載 · 252 次瀏覽
診斷 AI 代理中的上下文退化
長對話和大型提示可能會讓 AI 代理遺失重要資訊,或遵循互相衝突的上下文。此技能可協助 Claude、Codex 和 Claude Code 識別退化模式並選擇緩解策略。
建立可靠的 LLM 評估系統
LLM 評估很難在不同團隊與模型之間保持一致。此技能有助於設計評審提示詞、評分規準、指標與偏差檢查,以進行可重複的評估。
最佳化 AI 上下文視窗
長時間 AI 工作階段可能浪費詞元、拖慢回應速度,並遺失重要上下文。這項技能為 Claude、Codex 和 Claude Code 提供實用模式,用於壓縮、遮罩、快取和規劃上下文預算。
可靠地壓縮長代理脈絡
當脈絡被過度積極壓縮時,長時間的代理工作階段會遺失重要檔案、決策與後續步驟。這項 skill 提供結構化壓縮與評估方法,可在降低 token 使用量的同時保留可用脈絡。
最佳化代理程式上下文管理
當提示、工具、歷史記錄和檢索內容競爭上下文空間時,代理程式系統可能會變得不可靠。此技能說明上下文組成要素,以及用於預算配置、漸進式揭露和驗證的實用模式。
設計多代理架構
當單一情境視窗承載所有任務時,複雜的代理工作可能會失敗。此技能可協助 Claude、Codex 和 Claude Code 設計具備清楚交接流程的代理系統。
自動化 Grasshopper 工作流程
由 AmemiyaLai
Grasshopper 建模可能需要許多重複的元件、連線和參數步驟。此技能會將結構化設計筆記轉換為可執行的工作流程檔案和本機 MCP 動作。
Replace Text in Images with MakeItSay
由 makeitsay
Changing image text manually can require rebuilding fonts and backgrounds. This skill guides MakeItSay edits that replace or remove selected lines while preserving their appearance.
Split Flat Images into Editable PNG Layers
由 unmerge-app
Flat images combine subjects, text, and backgrounds, making separate edits difficult. This skill guides Unmerge processing to return positioned PNG layers with reconstructed hidden areas.
Plan McDonald's Meals by Nutrition and Budget
由 hessianz
Choosing meals within nutrition and budget limits requires checking menus, serving sizes, and discounts. Chilemai combines current MCP data with local meal planning and confirmed ordering.
Draft and Schedule Social Posts with TimeToPost
由 melmayssonowen
Managing platform-specific posts requires consistent wording, account selection, and approval. This skill guides TimeToPost MCP drafting, timing recommendations, and scheduling after explicit user approval.
Practice Notes and Transpose Chords with Clef
由 melmayssonowen
Music learners need focused help with note reading, chord symbols, and guitar tab. This skill combines guided drills, Clef MCP tools, and offline chord transposition.