Cheat Sheet

If you only read one page, read this one. Every claim below links to the deeper doc.

What is agentic coding (2 sentences)

Agentic coding means letting an AI agent — not just an assistant — read your project, run commands, and make multi-file edits on its own. The skill is giving the agent the right context, the right boundaries, and the right feedback loop. See What is Agentic Coding?.

The 3 tools most devs pick

Full comparison: Tool Comparison Matrix.

The 3 frameworks (for builders)

• Claude Agent SDK — Claude Code as a library, Python + TypeScript. Best default if you use Claude.
• OpenAI Agents SDK — lightweight, Python-first. Best default if you use OpenAI.
• Pydantic AI — Python, model-agnostic, integrates with Pydantic validation and Logfire.

Rule of thumb: if your workflow is a straight line (A → B → C), use plain async/await. Reach for a framework only when you need agent autonomy or branching. See Framework Comparison.

MCP in one paragraph

MCP (Model Context Protocol) is the "USB-C for AI agents" — an open standard that lets an agent call tools and read resources exposed by an MCP server. Servers run locally (stdio) or remotely (Streamable HTTP). Start with npx -y @modelcontextprotocol/server-filesystem /path for local files, or npx -y @playwright/mcp@latest for browser automation. Claude Code, Cursor, Copilot, Windsurf, and Zed all support MCP as of 2026. See MCP Guide.

CLAUDE.md in one paragraph

A CLAUDE.md at your project root tells Claude Code what your project is, what commands exist (pnpm dev, pnpm test), what code style to follow, and what gotchas to avoid. Keep it under 200 lines — every line competes for attention. Run /init once to auto-generate, then prune aggressively. Use .claude/settings.json for real permissions (not CLAUDE.md — that's advisory). See CLAUDE.md Best Practices.

The 5 patterns worth knowing

1. Structured output via tool use — force JSON via a tool schema, not prompting. Under 0.2% parse failures vs 8–15%.
2. Prompt caching — mark the long system prompt with cache_control: { type: "ephemeral" }. 90% off on cache hits.
3. Batch API — 50% discount for non-realtime workloads. Stacks with caching → up to 95% off.
4. Chain-of-thought for scoring — force a reasoning field before the numbers. Prevents clustering at 7/10.
5. Model routing — cheap model (Haiku) for classification, expensive (Sonnet/Opus) only for quality writing.

One thing that will bite you per tool

• Claude Code — CLAUDE.md is advisory. Use permissions.deny in .claude/settings.json for real restrictions.
• Cursor — .cursorrules (single file) is legacy. Use .cursor/rules/*.mdc with frontmatter for path-scoped rules.
• Copilot — gh copilot subcommand is retired; use the standalone GitHub Copilot CLI.
• Windsurf — Cascade flow state is per-session; it doesn't persist the way CLAUDE.md does across sessions.
• Codex CLI — Sandboxing is on by default. Lift explicitly per task.
• MCP — every active server adds tokens to your context. Cap at 5–6 servers.

One thing that will bite you in production

• Prompt injection — never trust tool-returned text (web content, file contents) as if it came from the user. See Security.
• Context window fills fast. Use /clear between unrelated tasks, use sub-agents for investigations. See Context Management.
• Unsandboxed file access. An agent with filesystem + shell access can brick your machine if a prompt injection succeeds. See Permissions.

Where to next

• New builder? → Choosing Your First Tool
• Framework picker? → Framework Comparison
• Going to production? → Cost Optimization + Evaluation + Observability
• API consumer? → API Reference
• Subscribing to digests? → Delivery Channels