AI Model Mentor

Turn your AI coding assistant into a cautious 10-year full-stack mentor — pure prompts, zero dependencies. Load this framework at the early stage of your project to make your AI act more cautiously and produce code that is easier to maintain. Pure prompt implementation, no dependencies.

中文一句话介绍：一个「纯提示词」框架，把你的 AI 编程助手变成一个有 10 年经验、谨慎靠谱的全栈导师，专为编程零基础小白设计。加载后 AI 会默认遵守 6 条铁律（代码即文档、安全前置、零破坏性改动、分步执行、模块化隔离、性能与资源前置），产出安全、可维护、带文档的代码。

🌍 Language / 语言

English · 中文 · 日本語 · 한국어 · Español · Français · Deutsch · Português · Русский

---

What is this?

A pure-prompt framework that shapes your AI coding assistant into a full-stack architect & development mentor with 10 years of experience, built for coding beginners with zero foundation.

It forces the AI to follow a set of "iron rules" — making Security First, Transparent Logic, Documentation First, Token Efficiency, Phased Implementation, and Resource Control its default behavior. The result: AI that doesn't just write code, but writes safe, maintainable, documented code.

✅ Tool-agnostic: works with opencode, Claude Code, Codex, Cursor, Gemini CLI, Google Jules, Aider, Windsurf, GitHub Copilot Agent — only the loading method differs (see the table below).

Core Modules

Module	File	Purpose
🧑‍🏫 Mentor Role	AGENTS.md	Architect-mentor persona + 6 iron rules + security & performance self-check checklist
🛡️ Security Spec	security.md	8 security domains: secrets / input validation / database / XSS / file system / external requests / error handling / performance & resource
🎨 Interaction Style	style.md	Life analogies, phase tags, confirm-before-act, progressive complexity
📋 Dev Workflow	workflow.md	Docs system / resource estimation / database design / frontend mapping protocol / deploy & rollback / test loop / version anchors

The 6 Iron Rules

1. Code as Documentation — all code carries comments explaining the "why"
2. Security Upfront — no hard-coded secrets, strict input validation, parameterized queries, XSS prevention
3. Zero-Destructive Changes — analyze dependencies first, tag edits as [Mandatory] / [Optional]
4. Phased Execution — never more than 300 lines per output, wait for confirmation at each step
5. Modular Isolation — max 500 lines per file, reserve extension interfaces
6. Performance & Resource Upfront — output index design with database schema, default pagination for list queries, three-tier resource estimation (memory/disk/CPU) at project start, release mechanisms for large memory operations

🛠️ Supported AI Tools

The prompt content is tool-agnostic — every tool only differs in how it loads files (filename, location, command). See zh-CN/COMPATIBILITY.md for the full per-tool guide (also available in all 9 languages).

Tool	Main file (agent)	Location	Loading
opencode	AGENTS.md	project root	auto
Claude Code	CLAUDE.md / AGENTS.md	project root	auto
OpenAI Codex	AGENTS.md	project root	auto
Cursor	AGENTS.md	.cursor/rules/	auto
Gemini CLI	GEMINI.md	project root	auto
Google Jules	JULES.md	project root	auto
Aider	CONVENTIONS.md	project root	auto
Windsurf	.windsurfrules	project root	auto
GitHub Copilot Agent	AGENTS.md	project root	auto
Any MCP client	via mentor-mcp	stdio (node mcp/dist/index.js)	auto (resources + tools)

💡 One-click install: the mentor CLI writes files to the correct name/location for any of these tools automatically (see cli/).

🧩 MCP Server (IDE)

mcp/ is an on-demand policy + sandbox MCP server. The IDE is the coding agent; this process does not run an LLM and is not on npmjs.

Users: download guapimm-mentor-mcp-*.tgz from GitHub Releases (self-contained; no tsc). Config: mcp/examples/mcp.release.json.

From source:

git clone https://github.com/guapimm/AI-Model-Development-Mentor.git
cd AI-Model-Development-Mentor/mcp
npm install          # local deps only — not a registry publish
npm run build

Copy mcp/examples/mcp.json, replace the placeholder with your absolute path. First tool call: session_start. Full tutorial: mcp/README.md.

Usage rules

1. Load the smallest set that fits the task. AGENTS.md is enough for everyday coding. Open security.md / workflow.md / style.md only when you need them. With MCP, call session_start then policy_load(id) — do not paste all four modules every turn.
2. Phase 0 before application code. Requirements + resource estimate first; then Design → Logic → UI → Test, confirm at each step.
3. Prefer MCP sandbox tools when the server is connected. fs_write / run_command enforce line limits, .env deny, and phase gates. The IDE’s own Write/Bash bypasses those gates.
4. Do not treat complete as the source of truth. It is an older one-shot dump and is missing later resource-control rules. Use the four split modules.
5. Secrets live in environment variables. Only .env.example (names, no values) belongs in git.

Notes / caveats

• Install paths: GitHub Release mentor binary (prompts only, no Node); GitHub Release guapimm-mentor-mcp-*.tgz (IDE MCP, Node ≥ 18, no local tsc); or clone + build mcp/. There is no npx @guapimm/mentor-mcp on npmjs.
• MCP config needs absolute paths and a reload after edits. Windows: E:/path/to/repo is fine.
• Sandbox is a path jail, not a VM. It keeps I/O inside the workspace and blocks .env / .git; Docker is optional for run_command only.
• Node ≥ 18 is required only for MCP. Prompt-file users can ignore mcp/.
• After git pull, rebuild: cd mcp && npm install && npm run build.

⬇️ Install

Two ways, same result:

# Option A — binary (zero dependencies, cross-platform)
# Go: mentor-*  |  Rust: mentor-rust-*  (both on GitHub Releases), then:
mentor install

# Option B — manual
# copy files from <lang>/prompts/ per your language's COMPATIBILITY.md
cp en-US/prompts/AGENTS.md AGENTS.md

Both support: interactive wizard (language → modules → tool), non-interactive flags (--lang zh-CN --modules agent,security --cli claude-code --dir ./proj), and add / remove / list / detect / pack.

Quick Start (3 steps)

# 1. (Optional) Bootstrap a new project scaffold: requirements doc + .env.example + docs/
mentor init --name "my-app" --goal "a personal finance tracker"

# 2. Copy the mentor role into your project (rename it)
cp en-US/prompts/AGENTS.md AGENTS.md

# 3. (Recommended) Add security / style / workflow specs too
cp en-US/prompts/security.md security.md
cp en-US/prompts/style.md style.md
cp en-US/prompts/workflow.md workflow.md

4. Launch your AI tool and say:

"I'm a complete beginner. Here is my Project Requirement Specification: project name ____, core goals ____, user roles ____, core workflows ____, data to persist ____. Start from Phase 0: Environment Setup & Tech Stack Selection + Resource Estimation and guide me step by step."

The AI will advance through "Design → Core Logic → UI → Testing", waiting for your confirmation at every stage.

📐 Fill-in Templates

Pre-made tables referenced by the workflow module, ready to copy into your project:

• templates/{lang}/resource_estimate_template.md — the Phase 0 《Project Resource Estimate》 table.
• templates/{lang}/ui_mapping_template.md — the frontend UI/element + event mapping table.

File Structure

AI_Model_Development_Mentor/
├── README.md            # English landing page + language switcher
├── LICENSE              # Apache-2.0 License
├── cli/                 # mentor CLI (Go)
├── rust/                # mentor CLI (Rust, zero-dependency mirror)
├── mcp/                 # mentor-mcp (stdio; GitHub Release .tgz, not npmjs)
│   ├── examples/mcp.json
│   ├── examples/mcp.release.json
│   └── policy/fragments.json   # on-demand prompt slices
├── templates/           # fill-in templates (resource estimate / UI mapping)
├── zh-CN/  en-US/  ja-JP/  ko-KR/  es-ES/  fr-FR/  de-DE/  pt-BR/  ru-RU/
└── <lang>/
    ├── README.md        # language entry + usage guide
    ├── COMPATIBILITY.md # per-tool loading instructions (the "one adapter file")
    └── prompts/         # tool-agnostic content (per language)
        ├── AGENTS.md    # mentor role ★ required
        ├── security.md  # security spec
        ├── style.md     # interaction style
        ├── workflow.md  # dev workflow
        └── <full>.md    # consolidated one-shot prompt

📦 New tools are supported by adding a row in each language's COMPATIBILITY.md — no per-tool directories needed.

🚀 Releasing (maintainers)

Tag a version to trigger the CI pipeline (.github/workflows/release.yml):

git tag v0.1.0 && git push origin v0.1.0

The workflow builds both implementations — Go binaries (windows/linux/darwin × amd64/arm64) and Rust binaries (mentor-rust-*) — and uploads them to the GitHub Release. No secrets or tokens required.

FAQ

Q: Do I need all 4 modules? A: No. AGENTS.md is the only must-have. Add security.md for stronger guardrails, style.md for a friendlier conversation experience.

Q: Is mentor-mcp on npm? A: Not on npmjs. Get guapimm-mentor-mcp-*.tgz from GitHub Releases, or clone and build mcp/. See mcp/README.md.

Q: Does this work with other AI products? A: Yes. The prompt content is tool-agnostic — every tool just loads it differently. See the table above or COMPATIBILITY.md for each tool's loading guide.

Q: Is this translated by machine? A: All language versions are reviewed translations of the same Chinese original. If you spot an issue, feel free to open an issue or PR.

License

Apache-2.0 © 2026 guapimm