PromptingCraft Guidance

com.promptingcraft由 promptingcraft.com 發布官方發布者v0.3.0更新於 Oct 29, 2026

Source-backed prompting and coding-agent guidance with provenance and freshness checks.

已驗證Streamable HTTP可網頁執行知識與記憶開發者工具

概覽

AI 產生的概覽

唯讀的遠端 MCP 伺服器，提供附來源與查核日期的官方提示詞與編碼代理指引紀錄，並支援來源新鮮度檢查。

可以做什麼
  • 依目標編碼代理、模型系列、適用範圍與主題搜尋精選指引紀錄
  • 依 id 取得單筆紀錄，含提供者、來源網址、適用性、限制、標籤、查核日期、快照雜湊與帳本提交
  • 檢查某個白名單官方文件頁面自查核快照以來是否變動，回傳未變、已更新、暫時無法取得或未知
適用情境
在為 Claude Code、Codex、Cursor 等編碼代理撰寫或審閱提示詞時，需要適用於該代理與模型的指引、可引用的來源，以及判斷來源頁面自查核後是否變動時使用。
執行需求
需要可連線的遠端 streamable HTTP 端點；未宣告帳號、憑證、環境變數或請求標頭。用戶端需能透過 HTTPS 連線至該伺服器，無需本機套件或執行環境。
安裝前請注意
工具為唯讀：僅列出、回傳與比對精選紀錄，除新鮮度檢查會抓取一個白名單頁面外，不會從網路抓取內容。伺服器要求只送出篩選值與識別碼，不要送出對話文字、程式碼、檔案路徑或憑證。標示為 PromptingCraft 方法論的紀錄是廠商自有方法，並非供應商建議；新鮮度檢查只說明頁面是否變動，不代表指引正確。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 PromptingCraft Guidance，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/com.promptingcraft%2Fguidance/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "promptingcraft-guidance": {
      "type": "http",
      "url": "https://mcp.promptingcraft.com/mcp"
    }
  }
}

README

PromptingCraft.com

Website: promptingcraft.com

The product is publicly named PromptingCraft.com. Since 0.3.0 the skill and plugin identifier is promptingcraft (formerly promptcraft); see the migration notes.

Turn a development discussion into a prompt a fresh coding agent can execute. PromptingCraft.com preserves the task contract, consults applicable official guidance, drafts, critiques once, revises when necessary, and checks that requirements and permissions survived. The assistant writing the prompt and the agent executing it may differ. Targets: Claude Code and OpenAI Codex; Cursor is supported as a target agent (see Use in Cursor). The standalone Skill on the current source tree has version 0.4.0. Its plugin package is a separate 0.4.0 upgrade candidate; the published ChatGPT plugin remains 0.3.0, and the candidate has not been published or approved as its replacement. The existing remote guidance MCP server reports version 0.3.0; it is a separate service.

The immutable v0.1.0 tag remains the historical source release; it is not replaced by the separate ChatGPT plugin. This branch contains unreleased guidance, compatibility and package changes. The latest contract-fidelity sample found 0/14 material drift after earlier diagnostic variance; it is not a guarantee. Native Codex invocations are verified. ChatGPT-native activation and downstream superiority remain unverified. Development CI archives do not replace v0.1.0 release assets.

When the original task is already suitable, PromptingCraft.com can return: "The original prompt is already suitable. No material revision is necessary." It preserves the original text; no-change still requires contract and source checks.

It helps produce a usable handoff; it cannot supply missing access, authority, repository evidence or guarantees of downstream correctness.

Canonical source: GitLab. A public GitHub mirror is also available for browsing and installation. GitLab CI is authoritative; GitHub Actions are not enabled. Contributions may be proposed on either platform; changes are integrated through the canonical GitLab project.

Launch news

PromptingCraft.com is published in the ChatGPT Plugin Directory, and its optional guidance MCP is listed in the official MCP Registry. Both published services are version 0.3.0; the standalone Skill in the current source tree is 0.4.0. Read the launch announcement for installation, capabilities and verification details.

Two ways to use it

Use PromptingCraft as a standalone Agent Skill, or connect its MCP server for structured access to official AI-developer guidance.

Standalone Agent SkillSkill + MCP server
What you installskills/promptingcraft/ (instructions and dated references)The plugin: the same skill plus mcp.json
HostsCodex, Claude Code and Cursor (local skill folders)Hosts that load the plugin's MCP server; verified in Codex
Prompt authoring (contract, draft, critique, final, no-change)YesYes, same skill
Official guidanceBundled curated references, dated; the host's own documentation search or browsing when availableAlso search_guidance and get_guidance: records with provenance and applicability
FreshnessLive only when the host actually reads an official page in that run; otherwise labeled dated cachedcheck_source_freshness compares one official page with its verified snapshot (cached up to 15 minutes)
NetworkNone requiredHTTPS to https://mcp.promptingcraft.com/mcp (fixed categories and ids only; see privacy)
If guidance retrieval failsUses bundled references and says soSays the server was unavailable, then uses other official sources or bundled references

The MCP server is optional. It supplies evidence, not authority: the skill still extracts the contract, drafts, critiques and finalizes, and the user's requirements and permissions always win. Neither mode guarantees real-time documentation: bundled references carry their retrieval dates, and a freshness check shows only whether a page changed since it was verified. Records of kind promptingcraft_methodology are PromptingCraft.com's own method, not provider recommendations. Server details: mcp/README.md and data boundary.

A small example

Ask: "Write a Codex task: Save shows success after an API error. Preserve the API shape; don't deploy."

Draft: "Fix saving and run tests."

Critique: the draft lost the symptom, API constraint and deployment boundary.

Final:

text
Fix Save so an unsuccessful API response is not reported as success. Preservethe API shape. Locate the existing save path and verify the requested failurebehavior. Do not deploy.
Optional suggestion: use the existing error-display pattern if one is available.

This is an original illustration, not proof of improved execution. See the worked examples and evaluation protocol.

Use in ChatGPT

The PromptingCraft.com plugin bundles the skill and declares one optional remote MCP server (the guidance server, no account or credentials). Build its local source once, register the real repository root, and install through supported Codex or ChatGPT Desktop plugin controls. See the exact local installation and ChatGPT procedure. After installation, start a new chat; explicitly select PromptingCraft.com with @ or ask “Prepare a Codex prompt for this task.” It uses the established discussion and returns the final execution prompt in one copyable code block.

Current OpenAI documentation supports plugins in Chat, but local-source and account availability vary. Codex CLI installation is verified separately from ChatGPT Desktop and ordinary Chat activation. Computer Use cannot access ChatGPT in this environment; those native checks remain a documented owner-side gate. No repeated SKILL.md paste or upload is part of the supported installation route. With unavailable retrieval, source guidance remains dated/unavailable.

Install the standalone skill

From the project where you want Claude Code to use the Skill, install the current public GitHub source with the skills CLI:

sh
DISABLE_TELEMETRY=1 npx -y [email protected] add revanbcn/promptingcraft-skill --skill promptingcraft --agent claude-code --copy --yes

This installs the standalone Skill from the repository's current default branch into .claude/skills/promptingcraft/ in the project. That branch carries Skill version 0.4.0; it can change as the repository is updated. The command opts out of CLI telemetry and uses the CLI's non-interactive copy mode. Installation does not establish Skill activation or install the optional remote MCP server. To use the Skill, start a new Claude Code session and invoke /promptingcraft with your prompt-authoring request.

For the historical release, download the v0.1.0 skill archive and SHA256SUMS from GitLab v0.1.0. Verify checksums before extracting. There is no 0.2.0 release yet; a clone provides the current candidate:

sh
git clone https://gitlab.com/revanalex/promptingcraft-skill.gitcd promptingcraft-skillpython3 scripts/install_skill.py --host codex --project-dir /path/to/your/project# Or, for Claude Code:python3 scripts/install_skill.py --host claude --project-dir /path/to/your/project# Or, for Cursor:python3 scripts/install_skill.py --host cursor --project-dir /path/to/your/project

The optional installer requires Python 3.11+, refuses existing targets and touches only the explicitly chosen project's skill directory. You can instead copy the extracted promptingcraft/ directory yourself into .agents/skills/ (Codex) or .claude/skills/ (Claude Code) or .cursor/skills/ (Cursor; Cursor also reads .agents/skills/), including all its references and license. Restart the host if discovery has not refreshed. In Codex invoke $promptingcraft; in Claude Code invoke /promptingcraft, then describe the prompt you need. No development task is executed by the skill.

Use in Cursor

Cursor reads Agent Skills from .cursor/skills/ or .agents/skills/ in a project and from ~/.cursor/skills/ for the user. The plugin root (plugin.json with the Agent Plugins 1.0.0 schema and skills/promptingcraft/) follows Cursor's documented Agent Plugins format; it has not been submitted to or reviewed by the Cursor Marketplace.

When the executor is Cursor, PromptingCraft.com loads its Cursor notes: under Auto the underlying model is treated as unknown, so no Claude, GPT or Composer prompting rules are applied unless the selected model is established.

Measured in the Cursor CLI (cursor activation eval): with the default Auto setting (reported as Composer 2.5 Fast) the skill was available but did not activate automatically (0/30 prompt requests); naming it, as in “Use the promptingcraft skill to prepare a Cursor prompt for this task”, loaded it 2/2 (measured with the 0.2.0 identifier and repeated with promptingcraft). In the editor you can also pick it from the / menu. Activation depends on the selected model. Plugin loading in the Cursor editor has not been tested.

Plugin package

The plugin ZIP includes root plugin.json using Agent Plugins 1.0.0, one skills/promptingcraft/ tree, listing icons/logos under assets/ and a Claude Code compatibility manifest, plus a root mcp.json that declares the optional guidance MCP server. It declares no hooks or credentials. Its ChatGPT listing metadata (display name, descriptions, website, support, privacy and terms URLs, brand colors and icons) is checked by the validator. The existing 0.3.0 plugin is published; this 0.4.0 candidate has not been submitted for review. In Claude Code, session-only loading is:

sh
claude --plugin-dir /path/to/extracted/promptingcraft-plugin# In that session: /promptingcraft:promptingcraft

The repository already includes .agents/plugins/marketplace.json for dondego-local. Prepare its actual plugin directory before registration:

sh
python3 scripts/prepare_marketplace.pypython3 scripts/validate.py --marketplace

The generated dist/promptingcraft-plugin/ comes from the canonical skill using the existing reproducible archive builder. It is a snapshot, not a second maintained skill. The catalog resolves its path relative to the repository root. See installation, refresh and removal.

Verification and freshness

sh
python3 -m unittest discover -s tests -vpython3 scripts/validate.pypython3 scripts/package.py --output distpython3 scripts/prepare_marketplace.pypython3 scripts/validate.py --marketplace# Optional live maintenance, kept separate from merge checks:python3 scripts/freshness.py --output /tmp/promptingcraft-freshness --write-baseline

Normal checks are deterministic, offline and use synthetic documents. Live source checks compare usable normalized content hashes, use HTTP validators when supplied and discover candidate links in official indexes. They report changed, unchanged, new, missing/unavailable and partial/error states. They neither prove guidance is correct nor automatically change the skill. Fetch time is not publication time. Caches and raw model traces stay outside tracked content; reports contain metadata, not wholesale provider documentation.

See freshness operation and schedule gate, verified source decisions, actual verification and the bounded roadmap. Source guidance is revalidated during skill use when host network capabilities permit; offline notes are never current.

Scope and license

MIT licensed original content. No official OpenAI/Anthropic affiliation or endorsement. No hosted app, telemetry, sign-up, model router, API billing, marketplace submission, deployment or automatic instruction rewriting in v0.1 or the 0.2.0 candidate. See LICENSE, CONTRIBUTING, SECURITY and CHANGELOG.

Developed by the dondego.com team.

來源：README.md，提交 7c57520

工具

3

search_guidance

Search curated guidance

List curated guidance records that match a target coding agent, model family, scope and topic. Returns official model guidance, platform documentation and Agent Skills standard records with labels, applicability and the date each source was verified. PromptingCraft.com methodology records are returned only when kind is promptingcraft_methodology and are never official provider guidance. Inputs are fixed categories only. Use get_guidance for the full record.

輸入參數kindlimitscopetopicmodel_familytarget_agent

get_guidance

Get a guidance record

Return one curated record by its stable id (from search_guidance): provider, canonical source URL, applicability, limitations, labels, the date the source was verified, its snapshot hash and the ledger commit. Unknown ids return an error; nothing is fetched from the web.

輸入參數id必填

check_source_freshness

Check source freshness

Fetch one allowlisted official documentation page from the curated registry and compare its normalized text with the verified snapshot. Status is unchanged, updated, unavailable (could not be checked now) or unknown (no snapshot). Results are shared by all callers and cached for up to 15 minutes (1 minute when unavailable); live.cached and checked_at show when the page was actually fetched. When the per-source fetch limit is reached, status is unavailable with reason rate_limited. Publication and update dates are never inferred and are always null. Only registry source ids are accepted; no other URL can be requested.

輸入參數source_id必填

版本歷史

1
  1. v0.3.0最新Oct 29, 2026

評分

已滿足適用檢查項的 100%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分
  • 已宣告認證方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分