Awesome AI Coding 技巧

★ 348Awesome Lists

#awesome#techniques#how-to

Overview

A community-maintained, practitioner-tested list of AI coding techniques (348 stars), organized by development stage — from requirements and planning through UI prototyping, coding, debugging, testing, review and refactoring. Each technique cites its source (Simon Willison, Armin Ronacher, Indragie Karunaratne, the Anthropic team, etc.) and community adoption rate, with concrete instructions for tools like Claude Code, Cursor and Codex CLI.

Who it's for

Anyone already using AI coding assistants who wants to turn scattered tips into a systematic methodology — especially Claude Code, Cursor and Codex CLI users.

What you'll learn

01Organization: arranged by development stage
The list is organized into seven stages: Requirements & Planning, UI & Prototyping, Coding, Debugging, Testing & QA, Review & Refactoring, and Cross-Stage Techniques. Each stage holds independent techniques with community adoption votes (e.g. 'Set Up Memory Files' 81% essential, 'Choose Boring, Stable Libraries' 64%) so beginners can see at a glance which practices are widely validated.
02Requirements and planning: build the foundation
This stage covers setting up memory files (using CLAUDE.md / AGENTS.md to give agents persistent project context and conventions), asking to plan first (using Plan Mode to stay read-only and list steps, risks and quick tests), planning with a high-capacity model, and spec-driven development (iterating on Markdown specs as the source of truth). The list quotes GitHub Engineering on treating specs as living documents refined to guide generation with increasing precision.
03Hands-on techniques for coding and debugging
The coding stage covers writing critical parts yourself and delegating the rest, treating the assistant like a digital intern with exact function signatures, choosing boring stable libraries, and priming with existing code. The debugging stage covers pivoting when the assistant repeatedly fails, logging everything so agents can self-diagnose, letting the tool change code / run tests / see failures / retry, and spawning subagents to double-check details.
04Testing and review: hold the last line
The list stresses that you absolutely cannot outsource testing, quoting Simon Willison: you only have a working system once you've seen it run. For review it recommends treating AI code like a colleague's pull request, asking the agent to review its own work, always checking the full diff (vibe coding can introduce unintended side effects), and having one agent write while a fresh session reviews.
05Cross-stage techniques and companion site
Cross-stage tips include one goal per session, clearing the context window with /clear frequently, choosing the right model and reasoning level per task, centralising memory files (using @-references or symlinks so CLAUDE.md / AGENTS.md point to one canonical doc), running multiple agents in parallel, and enabling autonomous mode for low-risk tasks. The repo pairs with the live site coding-with-ai.dev for searching, voting and contributing new techniques.

Practice tips

Do one thing this week: create a CLAUDE.md (or AGENTS.md) at your project root capturing common commands, tech stack and gotchas, then compare one session's difference.
Try plan-first: use Plan Mode so the tool lists steps, risks and quick tests, and only green-light implementation after you review.
Delegate one real refactor to the AI while keeping your own full-diff review step, practicing treating AI code as a pull request.
Clear context with /clear between unrelated tasks, or open a session declaring a single goal, and compare the quality of the output.