Definition

Using LLM-based tools (IDE agents, MCP, project rules) to write, review, and refactor software, with the human responsible for architecture, verification, and integration.

Source-Backed Themes

  • Project context: Centralize stack, environment, style, and conventions so the model starts closer to “second attempt” quality rather than guessing blind (AI 辅助开发探索).
  • Roles: Treat the engineer as orchestrator of agents, mentor reviewing outputs, and problem solver rather than primary typist (same source).
  • Quality and safety: Always verify; pay extra attention to state, performance, and security; use layered review (AI pre-pass, human on architecture and business, normal team bar) (如何提升 AI 代码质量).
  • Iteration: Expect a multi-pass loop; early outputs mainly teach the system what the task really is (如何提升 AI 代码质量).
  • Tooling: Browser automation via Playwright MCP from Cursor (Cursor Playwright MCP).
  • Language for prompts: Consistent English verbs for instructions when prompting (prompt vocabulary).
  • Further reading: Prompting guides and “harness engineering” links as bookmarks (prompt engineering reading list, harness links).

Relationship To This Wiki

  • An LLM Knowledge Base is one way to give the agent durable, structured context beyond a single chat.
  • AI Knowledge Bases cover upload-and-query style grounding; development workflows often combine both patterns.

Current Evidence In Repo

Synthesis

The through-line is governance: AI accelerates drafting, but ownership of correctness, security, and maintainability stays with the human. Strong shared context (rules files, wikis, tickets) reduces rework; explicit review and iteration expectations reduce surprise when the first patch is wrong.