leluong141996-dev

Model Identity Skill — AI skill for Claude Code

AI community

Model-agnostic system prompt that aims to keep LLMs honest about their identity, capabilities, and knowledge cutoff, and to use web search (when available) instead of guessing.

How to install Model Identity Skill

This entry records only its repository, not the path inside it, so there is no exact command to give. Open leluong141996-dev/Model-Identity-Skill and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Model Identity Skill does

Model-agnostic system prompt that aims to keep LLMs honest about their identity, capabilities, and knowledge cutoff, and to use web search (when available) instead of guessing. Includes integration guides for Claude/GPT/Gemini/local models, an eval harness, and preliminary results on Qwen3-4B.

Alternatives in AI

  • Contributing To Claude Skills Library — Thank you for your interest in contributing to the Claude Skills Library 5.3k ★
  • Osaurus — Own your AI. The native macOS harness for AI agents -- any model, persistent memory, autonomous execution, cry 5.1k ★
  • Phantom — An AI co-worker with its own computer 1.3k ★

README

model-identity-skill

A model-agnostic instruction set that aims to keep an LLM honest about three things:

  1. Identity: who it is and who made it. It never impersonates another AI.
  2. Capability: what it can actually do in the current context. It never claims a tool or action it doesn't have.
  3. Knowledge boundary: what it actually knows. It answers from trained knowledge, marks uncertainty, and searches the web (when a search tool is available) instead of guessing. If it has no search tool, it says so.

**Tiếng Việt:** Bộ instruction dùng cho mọi LLM, nhằm giúp model giữ đúng danh tính, khả năng và giới hạn kiến thức của mình. Đây là prompt, không phải cơ chế bảo đảm: hãy tự kiểm tra trên model của bạn (xem phần kết quả sơ bộ bên dưới). Model chỉ trả lời bằng kiến thức đã được huấn luyện. Khi không chắc chắn hoặc thông tin có thể đã cũ, model dùng công cụ search (nếu có) để trả lời từ nguồn đáng tin cậy. Nếu không có công cụ search, model nói rõ giới hạn thay vì bịa câu trả lời.

Preliminary results (read the caveats)

**This is not proof that the protocol works.** It is a small, early test by the protocol's own author. Treat it as a starting point, and **run the eval on your own model and your own prompts before relying on it.**

**What was tested:**

  • Model: one small model, Qwen3-4B (via Ollama).
  • Prompts: 19 test prompts. These are the 15 scenarios in `examples/eval-cases.md`, with 4 of them run both with and without a search tool.
  • Runs: each prompt was run 3 times per configuration.
  • Search: the search tool returned fixed, mocked results.
  • Grading: rule-based, then checked by hand.

Details are in [`eval/REPORT.md`](eval/REPORT.md).

**Pass rates** (57 runs per configuration):

Configuration Passed
No protocol 43/57
Deployment facts only (name, date), no rules 44/57
Full protocol 49/57
Compact protocol 50/57

**How to read this:**

  • **The differ