Designing a TypeScript harness for coding agents

Viewing the codebase as an execution environment for agents, it argues that fast verification, narrow blast radius, and pre-commit guardrails are what enable Claude Code or Cursor to produce maintainable code.

• It uses a pnpm-based monorepo split into apps/, packages/, and docs/, with eslint-config, typescript-config, shared-utils, and shared-types as shared packages.
• Each app imports the shared ESLint config to apply the same rules.
• Common skills live in .agents/skills, with .codex and .claude connected via symlinks so that multiple agents use the same best practices.
• CLAUDE.md contains the architecture, Notion documentation update rules, and AI coding rules, managing code and documentation together.
• It encapsulates team-wide conventions by referencing skills like nestjs-best-practices and typescript-advanced-types, and reduces slop with prohibition rules such as No casting, No narration comments, No single-use abstractions, and No speculative error handling.
• Every 24 hours, GitHub Actions runs Claude Code to clean up dead code and outdated documentation, and also keeps CLAUDE.md up to date.