AI Agent Workflow Import Kit: Building a Human-led Architecture for Human-AI Collaboration

Core Guide to AI Agent Workflow Import Kit

Original Author/Maintainer: gastercat, Source Platform: GitHub, Release Date: June 1, 2026. This open-source kit targets education and team collaboration scenarios, offering a systematic guide to importing AI Agent workflows. Its core is the Human-led Architecture, which emphasizes humans holding decision-making authority while AI acts as an assistant. It addresses challenges in AI usage (e.g., weakened learning processes, code quality and security issues) through auditable and traceable workflows, applicable to education, team collaboration, tool building, and other scenarios.

Background and Motivation

With the improvement of large language model capabilities, AI Agents (such as GitHub Copilot, Claude Code) have become important tools for work collaboration, but they also bring new challenges: How to ensure AI usage does not weaken human learning? How to maintain code quality and security? How to establish auditable and traceable collaboration processes?

The gastercat team developed this kit to address these challenges, providing a complete governance framework to help educational institutions and research teams import AI Agent workflows in a responsible manner.

Core Concepts of Human-led Architecture

This architecture emphasizes that humans always control requirement definition, architecture decisions, approval and review, and final responsibility, while AI serves as a restricted assistant. It includes four core principles:

1. Human-led Decision-making: Key decisions like project goals and technical selection are made by humans; AI only provides suggestions;
2. Auditability: Modifications involving AI must leave traceable records (Git version control, commit messages, code review);
3. Test-driven: AI-generated code must pass the same test standards as traditional code;
4. Progressive Authorization: From Inspect→Plan→Approve→Patch→Test→Review, each stage has clear human participation points.

Kit Structure and Content

It contains six core documents tailored to the needs of different roles:

• Professor's Summary: For teachers, explaining teaching objectives, learning value, risks, and evaluation directions;
• Main Document: Defines the overall architecture, core concepts, tool positioning, and import process;
• Student's User Guide: Guides students on using AI Agents, including Git basics, branching strategies, and code review norms;
• Building Guide: For developers/teaching assistants, covering tool installation, local model deployment (Ollama), and test environment setup;
• XML Prompt Templates: Structured prompts that limit task scope and define input/output formats;
• Risk Management and Scoring Recommendations: Provides a risk assessment framework, Git diff review guide, and scoring standards.

Recommended Import Process

The kit defines a seven-step process to ensure AI usage is under human supervision:

1. Inspect: AI first checks the codebase, understands requirements, and builds context;
2. Plan: AI proposes a modification plan, including impact scope, verification methods, and risks;
3. Approve: Humans review and decide whether to approve the plan;
4. Patch: AI makes modifications within the approved scope, isolated via branches;
5. Test: Execute tests to verify the correctness of modifications;
6. Review: Humans read Git diff to confirm the rationality of modifications;
7. Commit: Humans organize commit messages and formally submit the changes.

XML Semantic Constraints and Invariants Rules

XML Semantic Constraints: Use structured XML prompts to clearly distinguish tasks, scope, input/output, and forbidden items, reducing ambiguity. Templates include Inspect-only, Plan-only, Patch, etc.;

Invariants: Unbreakable rules in the project, such as API contracts, data formats, permission restrictions, and performance requirements, to prevent AI from breaking key constraints during modifications.

Applicable Scenarios and Target Audience

• Education Scenarios: Help teachers design courses for AI tool usage and establish evaluation standards;
• Team Collaboration: Define collaboration rules, branching processes, and code review standards for development teams;
• Tool Building: Guide developers to set up AI Agent environments (local model deployment, MCP integration);
• Personal Learning: Help individual developers build good AI usage habits and maintain code quality and learning effects.

Summary and Outlook

This kit represents a responsible attitude towards technology adoption. In today's era of rapid AI development, establishing clear usage boundaries and governance frameworks is crucial.

Human-led Architecture is not just a technical architecture but a work philosophy: it acknowledges the value of AI, insists on human-led decision-making, and achieves a balance between efficiency improvement and quality/transparency through auditable processes, structured prompts, and rule constraints.

In the future, such governance frameworks will become even more important, providing a starting point for sustainable collaboration models for educational institutions and teams.