Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Managing Knowledge for AI Coding Agents Across Multiple Repositories

AI summary

A team describes how it uses Claude Code and Codex with a main repository and more than 20 business repositories. Its development workflow divides knowledge into K0, K1, and K2 loading tiers; keeps business specifications alongside the relevant code; and gives target-repository worktree rules precedence over business-line and main-repository defaults. Formal knowledge is not updated during development, but reviewed and written back after delivery verification. When delegating to subagents, the team requires the task definition, task ID, target worktree path, rule-file paths, and the text and source of applicable K0 rules to be passed explicitly. The article presents these practices as the team’s own operating model.

In-depth insight

The account provides a concrete model for controlling coding-agent context, resolving rule conflicts, passing rules to subagents, and keeping formal business knowledge synchronized with verified code changes in a multi-repository workflow.

规则和业务知识全塞进一个文件会越塞越长、没人维护;我们的三条做法:按怎么读分三层、业务知识放回代码仓、开发中不改正式知识

这个系列记录我们团队怎么让 Claude Code 和 Codex 参与整个研发流程。开发阶段拆成三篇:上一篇(AI-SDLC 实践 05)讲了多仓库的工程结构——用一个主仓登记二十多个业务仓库,Agent 统一在主仓启动,代码则写在每个业务仓库从基线切出来的 worktree 里。这一篇讲开发阶段的知识库,下一篇讲 D2C。

让 Agent 真正进项目写代码,光看得懂语法不够,它至少得知道两类东西:一是团队的编码规范和架构约定,二是当前这个页面、这个模块的业务规则。

最常见的起步做法,是把能想到的要求全塞进 CLAUDE.md 或 AGENTS.md。但这套做法很快就会撞墙:文件越塞越长,上下文被大量无关规则挤占;更麻烦的是,真实的业务规则散落在文档平台、二十多个代码仓和老员工脑子里,根本塞不进单个引导文件,塞进去了也没人维护。

#Claude Code#Codex

Original · 20 min readRead the original
Group chat

No comments yet. Start the conversation.