Extension Path: For Developers

繁體中文 | 简体中文 | English

🚀 First time installing Claude Code or writing CLAUDE.md / SKILL.md? The quick setup guide is resources/setup-guide.en.md D-E. Skip it if you already know this.

← Back to main path README · Continue here after Track A's A3 or Track B's Stage 7. Apply agentic AI to coding workflows.

Use Cases (Developer Scenarios × How AI Helps)

The table below splits a developer's day into 7 common scenarios. Each has a different pain point, and each calls for a different level of AI tooling:

Scenario	Pain point	How AI helps	Recommended tools (light → heavy)
AI pair programming	You forget syntax mid-flow or cannot recall a method name	Autocomplete + rewrite + explanation	Cursor / Copilot → Claude Code
Multi-file refactoring	Changing one class risks missed references; cross-file rename is error-prone	Batch refactors while keeping style consistent across many files	Cursor → Claude Code → codex-delegate
Code review (your own PR)	Reviewing your own diff makes it easy to miss problems	Find bugs / smells and check edge cases	Claude Code / Cline → Continue (CI)
Writing tests	TDD cases are easy to miss; coverage falls short	Generate pytest cases from signatures / specs	Claude Code + Aider
Debugging	Logs are thin; stack traces are hard to interpret	Explain traces, generate hypotheses, run minimal repros	Claude Code
Docs	Docstrings / READMEs lag behind refactors	Generate docs from code and update docs alongside PRs	Claude Code
CI / team automation	Manual review is repetitive; style varies across people	Run automated review / lint in GitHub Actions	Claude Code Action + Continue

💡 Individual vs team: the first 6 rows are personal daily workflows. The final row (CI) is team governance. For teams under 5 people, AI automation in CI often has low ROI; you can defer it.

Curated Projects

CLI agent comparison: 7 major CLI agents (Claude Code / Codex / OpenCode / Gemini CLI / goose / Aider / Hermes Agent) compared side-by-side in resources/cli-agents-guide.en.md. New to CLI agents and want step-by-step onboarding → tracks/cli/A1-cli-intro.en.md (Track A first stop).

MCP catalog: Looking for integrations to wire CLI into daily tools (GitHub, Linear, Atlassian, Postgres, Playwright, Figma…) → resources/mcp-skills-catalog.en.md (62 entries by category).

This page only lists tool entry points directly relevant to developer workflows.

Coding Agents

Cursor ⭐⭐⭐⭐⭐

Editor-integrated AI pair-programming tool. Widely adopted in AI editor tools and a useful baseline for comparing other IDE agents.

Aider-AI/aider ⭐⭐⭐⭐⭐

★ 44k+ · Apache-2.0 — git-aware CLI pair-programmer. Edits files in your repo directly and writes commits for you. The open-source reference for "git-native AI editing." Model-agnostic.

anthropics/claude-code ⭐⭐⭐⭐⭐

★ 120k+ — Anthropic's official agentic coding assistant. Skills + plugins ecosystem.

cline/cline ⭐⭐⭐⭐⭐

★ 61k+ · Apache-2.0 — VS Code extension, autonomous in-IDE agent: tool use, browser, step-by-step approval. The first pick for VS Code users wanting IDE-native agentic dev.

continuedev/continue ⭐⭐⭐⭐

★ 33k+ · Apache-2.0 — source-controlled AI checks, enforceable in CI. Represents the team / governance angle on coding agents.

OpenHands (formerly OpenDevin) ⭐⭐⭐⭐

★ 72k+ · MIT — open-source autonomous software development agent. More aggressive design than Aider / Claude Code — agent runs in its own sandbox and commits autonomously. Best for "throw a whole issue at it" scenarios.

block/goose ⭐⭐⭐⭐

★ 43k+ · Apache-2.0 — Open-source, extensible AI agent that goes beyond code suggestions — install / execute / edit / test, with any LLM. Supports multiple LLM providers and MCP, ships as desktop app, CLI, and API. (Repo now resolves to aaif-goose/goose.)

RooCodeInc/Roo-Code ⭐⭐⭐⭐

★ 23k+ · Apache-2.0 — VS Code coding agent with a "team of specialized modes" model. Different from Cline's single-agent flow.

Code Review

obra/superpowers ⭐⭐⭐⭐

20+ battle-tested skills including TDD patterns, debugging, collaboration patterns. Good source for code-review skill design.

Recommended Tools

• yamadashy/repomix ⭐⭐⭐⭐⭐ ★ 24k+ — Typical developer use case: package the whole codebase for a reviewer / refactor agent. Outputs a single AI-friendly file (XML / Markdown / JSON) for Claude Code / Codex code review / refactoring. See the official README for technical details such as MCP server mode, tree-sitter compression, and secretlint filtering. A must-have, daily-driver-grade tool for Track A.

Workflows to Master (by frequency)

Frequency	Workflow	Steps (≤3)	Recommended tools	Best for
Daily	AI pair programming	(1) Open a branch (2) Give the task to Claude Code and ask for a plan first (no code yet) (3) Review plan → approve → code → review your own diff	Claude Code / Cursor / Cline	All developers
Daily	Git-native AI editing	(1) aider (2) Ask in natural language (3) review + commit / /undo	Aider	People who want a clean git flow
Per PR	Automated code review	(1) .github/workflows/claude-review.yml (2) Capture git diff → run prompt → post back to PR (3) human + AI review	Claude Code Action + Continue	Teams
Per feature	Test generation	(1) Provide function signature + docstring (2) Ask AI for pytest cases, including edge cases (3) Run coverage + intentionally break a bug to verify tests catch it	Claude Code / Aider	Test-writing phase
Occasional	Multi-file batch edits	(1) Claude writes a plan (2) codex-delegate handles mechanical refactors (3) Claude reviews the diff	Claude + codex-delegate	Refactors across 30+ files

💡 Starter habit: run "daily AI pairing" and "test generation" for a month first, then add automated PR review.

3 Concrete Workflow Recipes

1. AI Pair Programming (daily cadence) 1. Start a feature → git checkout -b feature/xxx 2. Hand the task to Claude Code / Cursor — make it write a plan first (don't dive into code) 3. Review the plan, course-correct → only then approve coding 4. After it's done: run tests + lint → review the diff yourself (don't blind-accept) 5. Write the commit message yourself, or have AI draft and edit before committing

2. Aider Git-Native Flow (closest "pair with AI" experience)

# Inside the repo
aider --model anthropic/claude-sonnet-4-20250514

# Natural-language ask
> Add a timezone parameter to parse_date in utils.py, default UTC

# Aider edits + commits automatically. To roll back:
> /undo # undoes the last AI commit

3. PR-time Claude code review (GitHub Action)

.github/workflows/claude-review.yml:

on:
  pull_request:
jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - name: Run Claude review
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          # Use anthropics/claude-code-action or your own script
          # Get git diff, run prompt, post results back to PR

Reference: official anthropics/claude-code-action GitHub Action.

Common Pitfalls (Anti-patterns)

❌ Don't	✅ Do instead
Let AI push directly to main	Always go through PR → review → merge
Blind-accept large refactor diffs	Break into < 50 LOC chunks, review each
Hand .env / API keys to the AI	Use your tool's exclusion mechanism — Cursor .cursorignore / Aider .aiderignore / Claude Code permissions.deny in .claude/settings.json
Let AI run shell freely against production code	Sandbox + permission whitelist
Take AI-generated tests at face value	Run coverage + intentionally break a unit to see if tests catch it
Discover wrong direction after many commits	Plan-first mode: review the plan before any coding

Tier Progression

Recommended progression:

Tier	Tools	Best for	Learning cost
Tier 0	Cursor / Copilot / Claude.ai	IDE chat, autocomplete, no custom agents	0 (if you can use an editor)
Tier 1	Claude Code / Cline / OpenCode + CLAUDE.md	CLI with file-system access, human-in-the-loop	1-2 days
Tier 2	Custom Skills + MCP server	Packaging dev workflows as shared team skills	1 week of setup
Tier 3	Auto-running agents in CI + production observability	Stage 7 territory	Several weeks, governance required

Most individual developers can stay at Tier 0-1. Validate ROI before going Tier 2+: it is only worth the investment if the team is large, the workflows repeat often, and failures are hard to reverse.

Other Branches Also Apply

Branches that overlap heavily with developers:

• Doing ML research / writing papers → Researcher branch
• Wire Notion / Linear / Atlassian / Postgres / Figma into your CLI → resources/mcp-skills-catalog.en.md
• Author your own Skill / MCP server → Stage 5 + resources/cookbook.en.md
• Schema design details → resources/schema-design-cheatsheet.en.md
• CLI from zero → Track A (A1 → A2 → A3)

Community Note

Contributions especially welcome:

• IDE-specific config templates (Cursor .cursorrules, Claude Code CLAUDE.md for Python / Go / Rust, etc.)
• Language-specific Skills (Python / TypeScript / Rust / Go best-practice patterns)
• CI / pre-commit hook integration case studies
• Multi-developer team governance — sharing Skills across devs, permission design, cost tracking

See CONTRIBUTING.md.