AI-Assisted Development
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
- OpenClaw agent + memory-wiki
- AI 辅助开发探索
- 如何提升 AI 代码质量
- Cursor Playwright MCP
- Prompt vocabulary
- Prompt engineering reading list
- Harness engineering links
- Canace blog index (catalog of longer posts, many on AI coding)
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.