cr-730/agent-system-prompt-architect-skill

★ 11PythonAudience · developerComplexity · 2/5ActiveLicenseSetup · easy

Mindmap

mindmap
  root((agent-system-prompt-architect-skill))
    Inputs
      Agent description
      Task scope
      Tool inventory
    Outputs
      Structured system prompt
      Role and non goals
      Format examples
    Use Cases
      Build agent prompts
      Avoid common prompt mistakes
      Evaluate prompt quality
    Tech Stack
      Codex
      Markdown
      Python
      Skills

mindmap root((agent-system-prompt-architect-skill)) Inputs Agent description Task scope Tool inventory Outputs Structured system prompt Role and non goals Format examples Use Cases Build agent prompts Avoid common prompt mistakes Evaluate prompt quality Tech Stack Codex Markdown Python Skills

Click or tap to explore — scroll the page freely

Why anyone would actually build this

REASON 1

Generate a structured system prompt for another agent with role, task scope, non goals, and failure handling

REASON 2

Avoid common mistakes like inventing tool names or writing only negative rules

REASON 3

Add few shot examples, reasoning policy, and retrieval grounding to an agent prompt when needed

REASON 4

Measure prompt quality against the bundled evaluation baseline

Stack

MarkdownPythonCodex

Spinning it up

Difficulty · easy Time to first run · 5min

Install is copying the skill folder into ~/.codex/skills and restarting Codex.

MIT license, so you can copy, modify, and use the skill commercially as long as you keep the copyright notice.

Wtf does this do

This repository is a skill for the Codex coding agent and similar AI code agents. The author explains, in Chinese, that they kept asking Codex to build agent style projects, and Codex kept producing system prompts that were not well structured. This skill teaches the agent how to write a high quality system prompt for another agent it is building. The README is clear about what the skill is for and what it is not. The right use case is when you ask the agent to write the system prompt for another agent. Wrong use cases include writing a casual one line prompt, marketing copy, articles, scripts, or general chatbot prompts. The README also lists the common mistakes the skill is designed to fix: treating the project name or brand name as a role, copying backend code identifiers into the system prompt, writing only negative do not rules with no positive constraints, inventing tool names and parameters that do not exist, and producing long unstructured walls of text. The skill produces a system prompt you can paste straight into an agent, with a clear role, task scope, non goals, and failure handling. It distinguishes semantic abilities from real runtime tool specs, covers sources, retrieval, citation, conflicting evidence, and uncertainty, and applies basic prompt engineering rules such as concrete instructions, format examples, and measurable success criteria. Advanced techniques like few shot examples, reasoning policy, ReAct style tool use, and retrieval grounding can be added when needed. Installation copies the skills slash agent-system-prompt-architect folder into your tilde slash .codex slash skills directory and restarts Codex. The README shows how to invoke the skill inside Codex by prefixing the request with dollar agent-system-prompt-architect followed by a description of the agent you want built. The repository also includes an evaluation suite with a baseline showing a 0.917 quality pass rate when the skill is used versus 0.554 without it. License is MIT.

Yoink these prompts

Prompt 1

Copy the agent-system-prompt-architect folder into ~/.codex/skills and restart Codex to load it

Prompt 2

Inside Codex, invoke the skill with $agent-system-prompt-architect and ask it to write a system prompt for a customer support agent

Prompt 3

Show me the list of common system prompt mistakes this skill is designed to fix

Prompt 4

Run the evaluation suite and compare the 0.917 quality pass rate with the 0.554 baseline

Prompt 5

Extend the generated system prompt with a ReAct style tool use policy and a retrieval grounding section

View the repo → Decode another repo

← cr-730 on gitmyhub — every repo by this author, as a profile.

Don't trust strangers blindly. Verify against the repo.