R&D Management 11 min read

Enterprise AI Team Development Guide: Tiered Process to Balance Speed and Quality

This guide introduces a tiered workflow for AI development teams—layered reading, graded intervention, and six ready‑to‑use AGENTS.md templates—mapping project types and demand sizes to lightweight, standard, or strict processes, enabling fast handling of simple tasks while ensuring stability for complex work.

Tech Verticals & Horizontals
Tech Verticals & Horizontals
Tech Verticals & Horizontals
Enterprise AI Team Development Guide: Tiered Process to Balance Speed and Quality

Problem Statement

When using the Codex + OpenSpec + Superpowers stack, teams face a binary choice: run every change through the full heavyweight workflow (which cuts efficiency roughly in half) or skip the workflow entirely (which leads to uncontrolled code quality, compatibility issues, and hidden risk).

Goal

Provide a concrete, tiered process that automatically matches workflow intensity to demand complexity while keeping the habit “read → decide → act” consistent across the organization.

Core Concept

Layered reading → graded intervention → template execution . Different project types and demand sizes use different paths instead of a one‑size‑fits‑all process.

1. Layered Reading – What to Read First

Reading order varies by project type.

New (Greenfield) Project

README – understand the purpose.

AGENTS.md – learn the team rules.

openspec/specs – basic behavior definitions.

Project structure – module layout.

Current task details.

Existing (Brownfield) Project

AGENTS.md – rule set.

Relevant module code.

Test cases.

Interface and call chain.

Historical specs (if any).

Simple Project

Prioritise speed; avoid over‑engineering.

Complex Project

Control risk first; analyse system, dependencies, and overall architecture.

Small Demand

Determine whether the change alters behavior. If not, write code directly; if it does, add a minimal spec.

Large Demand

Follow a fixed path: Current state → Constraints → Solution → Tasks → Implementation → Verification .

2. Graded Intervention – Matching Process Strength to Demand Size

The workflow is split into three levels.

Level 1 – Lightweight Mode

Applicable to small tweaks, ordinary bug fixes, and simple projects.

Combination: Codex + Simple Plan .

Level 2 – Standard Mode

For regular business requirements, interface changes, and multi‑module impact.

Combination: Codex + Superpowers (plan) + OpenSpec (minimal set) .

Level 3 – Strict Mode

For core system modifications, architectural adjustments, and cross‑service demands.

Combination: Codex + Superpowers (full process) + OpenSpec (complete documentation) .

3. Template Execution – Six Ready‑to‑Copy AGENTS.md Files

Each template targets a specific scenario.

1. New Project (Greenfield)

# AGENTS.md(新项目)
## 项目特点
新建系统,无历史包袱

## 核心原则
先设计结构,再开发

## 执行流程
1. 明确领域边界
2. 创建基础 spec
3. 输出实现计划
4. 分步开发
5. 验证

## 要求
- 核心功能必须先有 spec
- 避免过度设计
- 保持结构清晰

2. Existing Project (Brownfield)

# AGENTS.md(老项目)
## 项目特点
已有系统,需保证稳定性

## 核心原则
不破坏现有行为

## 执行流程
1. 阅读代码与测试
2. 分析影响范围
3. 再决定实现方式

## 要求
- 优先保持兼容
- 避免无关改动
- 优先增量改造

3. Simple Project

# AGENTS.md(简单项目)
## 核心原则
流程最简化

## 执行流程
1. 理解需求
2. 简要计划
3. 实现
4. 验证

## 要求
- 不引入复杂流程
- 保持代码简单

4. Complex Project

# AGENTS.md(复杂项目)
## 核心原则
风险优先控制

## 执行流程
1. 分析系统
2. 识别依赖
3. 设计方案
4. 输出 spec
5. 分步实施
6. 验证

## 要求
- 禁止大规模一次性修改
- 必须拆分任务

5. Small Demand

# AGENTS.md(小需求)
## 核心原则
快速完成

## 执行流程
1. 阅读相关代码
2. 简要计划
3. 实现
4. 验证

## 要求
- 最小改动
- 不影响其他模块

6. Large Demand

# AGENTS.md(大需求)
## 核心原则
可控推进

## 执行流程
1. 分析需求
2. 设计方案
3. 创建 spec
4. 拆解任务
5. 分步实现
6. 验证
7. 归档

## 要求
- 必须使用完整流程
- 必须考虑风险与回滚

4. Practical Adoption in Teams

Two‑layer usage : a global AGENTS.md defines company‑wide principles; a project‑specific AGENTS.md defines service‑level rules.

Template selection : pick one main template per project; use the graded levels for different demands within the same project; avoid frequent template switching.

Guideline : execute simple tasks quickly, handle complex tasks with the stricter levels, and never apply a single process to all work.

5. Supporting Skill Library

A ready‑to‑use skill library is available at

https://github.com/SYZ-Coder/superpowers-openspec-team-skills

. It contains copy‑ready SKILL.md workflows, lightweight proxy metadata, and usage guides that integrate with Codex‑compatible runtimes and the tiered process.

Key workflows in the library:

OpenSpec + Superpowers full workflow (requirement clarification → verification).

Superpowers → OpenSpec → Superpowers (explore, formalise, execute).

Superpowers‑only workflow (design, plan, TDD verification) – matches Level 1.

OpenSpec‑only workflow (spec first, then implement) – matches Level 2/3.

Quick start steps:

Read the library README (team-skills/README.md).

Follow the installation guide (team-skills/INSTALL.md).

Copy the desired workflow folder into the skill directory (e.g., .codex/skills/).

Invoke in a Codex prompt, e.g.,

Use $openspec-superpowers-workflow to run this feature from clarification through verification.
AI developmentCodexR&D processOpenSpecSuperpowersworkflow templates
Tech Verticals & Horizontals
Written by

Tech Verticals & Horizontals

We focus on the vertical and horizontal integration of technology systems: • Deep dive vertically – dissect core principles of Java backend and system architecture • Expand horizontally – blend AI engineering and project management in cross‑disciplinary practice • Thoughtful discourse – provide reusable decision‑making frameworks and deep insights.

0 followers
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.