CLI agent Introduced in L5

Claude Code

Anthropic's CLI coding agent. Runs in your terminal, in your repo, with full project context.

Mindmap

hover · click to navigate
this tech depends on / used by alternative Shipyard anchor
What it is

The plain-English version

Claude Code is a command-line coding agent from Anthropic. You run it in your repo; it reads files, runs shell commands, edits code, runs tests, all in conversation. Persistent memory via CLAUDE.md at the repo root, plus subagents, MCP server integration, and headless modes for automation.

Why it exists

The problem it solves

Different agents fit different work. CLI agents shine at multi-file, multi-step tasks where you want the agent driving for a while — refactors, migrations, test writing, debugging. The terminal context is right for engineers who already live there.

What it competes with

Alternatives

AlternativeTypeWhen it wins
CursorIDE agentAn AI-first IDE — VS Code fork with deep agent integration. Edit-with-AI, tab completions, repo-aware chat.
Where it shows up in Shipyard

Deep links

Vocabulary

The words you'll hear

CLAUDE.md
Project-level instructions and context. Lives at repo root, picked up automatically.
Subagent
A spawned helper agent for a specific task, with isolated context.
MCP server
Model Context Protocol — plug-in tools the agent can call (databases, APIs, etc.).
Headless / unattended mode
Run without interactive prompts. Useful in CI.
Plan / Apply
Some workflows show the plan before edits. Always read before applying.
Prompting

Bad vs. good prompt for Claude Code

✕ Bad prompt
fix my code
✓ Good prompt
Run the test suite (npm test) and look at the failing tests in src/server/api.test.ts. Read api.ts and api.test.ts. Hypothesize root cause first; don't edit. Once you have a hypothesis, propose the smallest fix and walk me through it before editing. If the fix touches more than two files or removes a test, stop and ask me first.

Why it works: Names the workflow (run, read, hypothesize, propose, edit), bounds blast radius (file count, test integrity), and asks for explicit approval gates. The agent stays useful and you stay in control.

Pitfalls

What bites real teams

⚠ No CLAUDE.md

Without project memory, the agent re-learns conventions every session. Capture them once.

⚠ Open-ended tasks

"Make this better" goes badly. Specify the deliverable, the scope, the success check.

⚠ Skipping diff review

The agent will tell you what it changed. Read the diff before committing — every time.

References

Official docs only