Anthropic coding agent
Claude Code is the agent this whole setup assumes you have.
Claude Code is Anthropic's coding agent for the terminal, editor, and desktop, and it is the default agent Groundwork installs. This page maps its surfaces, the daily commands, and the gauges that keep long sessions healthy.
#Which Claude surface to use
| Claude Code CLI | The daily driver. Start it in a project folder, give it a task, approve narrowly, review the diff. Installed by Groundwork (the claude-code@latest cask). |
| VS Code extension | The same agent attached to a visible editor buffer: inline diffs, @-mentions, and plan review without leaving the file. Groundwork installs it alongside VS Code for when you want the diff in a GUI. |
| JetBrains plugin | The same agent inside IntelliJ IDEA, PyCharm, WebStorm, and other JetBrains IDEs, running in the IDE terminal. Not installed by Groundwork; add it from the JetBrains Marketplace if that's your editor. |
| Claude desktop app | Anthropic's desktop app: general chat plus a Code tab for running and reviewing Claude Code sessions side by side. Groundwork installs it on personal machines only — work machines skip AI apps; reach for it when juggling threads outgrows the terminal. |
| claude.ai web | Chat, research, and cloud Claude Code sessions (claude.ai/code) from any browser, no install required — useful for a long task on a repo you don't have checked out locally. |
| Agent SDK | For building your own agents on top of Claude Code's tools and orchestration, in Python or TypeScript. Skip this unless you're building an agent product rather than using one. |
Use the CLI for daily repo work, an editor extension when you want Claude inside your editor (inline diffs in VS Code; the IDE terminal in JetBrains), and the desktop app or claude.ai web when a task benefits from parallel sessions or runs better off your laptop. Automated review on every pull request (GitHub Actions, GitLab CI/CD) is a separate, opt-in layer on top of the same agent — see AI agents & rules for the shared contract every surface follows.
#Daily commands
claude | Open the interactive session in the current project. |
claude "explain this repo" | Start with a one-shot prompt from the shell. |
claude --continue | Reopen the most recent session for this directory; --resume (or /resume) opens a picker across all of them. |
ShiftTab or --permission-mode <mode> | Cycle or set how much Claude can do without asking: Manual (the default — Claude asks before each edit or command; nothing changes without your yes) → Accept edits → Plan (read-only research). Auto and Bypass permissions are opt-in: Bypass skips every approval prompt, so reserve it for a throwaway container or worktree — never a machine holding credentials or work you can't lose. Trade approval friction for autonomy on purpose, not by accident. |
/plan <description> | Enter plan mode directly: Claude researches and proposes a plan without touching files, then asks how you want it to proceed. |
/context & /usage | The two gauges: context-window usage and plan limits. Check both before long work — see Know your AI budget. |
/compact <focus> | Summarize the session deliberately, telling the model what to keep. |
/rewind | Roll the conversation and/or the code back to an earlier checkpoint when a session goes sideways. |
/review [PR] | A fast, read-only, single-pass review of a GitHub pull request — findings only, no edits. |
/statusline | Put an always-on readout (model, context, whatever you script) at the bottom of every session. |
/model | Pick the model for the session — see "Which model to pick" below. |
Try it now: work a real change in plan mode first
In a project you know, run claude and press ShiftTab until the status line shows plan mode (or start with claude --permission-mode plan). Ask for a real change. Read the plan before you approve it — that's where you catch a wrong assumption before Claude touches a file.
#Which model to pick
/model switches the model for the current session and saves it as your default. The choice trades quality for speed and cost — see Know your AI budget for the budget side of that tradeoff.
| Claude Fable 5 | Anthropic's most capable model, tuned for long-running agents. Reach for it on the hardest design, architecture, or review work, or a task you expect to run unattended for a long stretch. |
| Claude Opus 4.8 | Built for complex agentic coding and enterprise work — a strong default for anything you'd call "the hard part." |
| Claude Sonnet 5 | The best combination of speed and intelligence, and Claude Code's default model. Good for most day-to-day work. |
| Claude Haiku 4.5 | The fastest model with near-frontier intelligence. Use it for mechanical work: renames, boilerplate, formatting, quick lookups. |
Rule of thumb: spend a heavier model on decisions you'd want a senior engineer's judgment on — design, security-sensitive changes, review — and let a lighter model burn through mechanical work where a wrong guess is cheap to catch and fix.
#The contract is the same as every agent's
Claude Code reads this repo's AGENTS.md through a thin adapter, inspects before editing, and is expected to verify before claiming done — the shared contract AI agents & rules defines for every tool. What differs per tool is the surface, not the discipline: inspect, plan, edit, verify, review, commit.
#Good first prompts
AGENTS.md, inspect this repo, and summarize the project structure and validation commands. Do not edit files yet.#Official references
- Claude Code quickstart — start here if you have never run Claude Code: install it and walk through a first real task.
- Claude Code docs — the full docs hub: every surface, workflow, and setting, once you're past the basics.
- CLI reference — every command and flag in one place, for scripting Claude Code or tuning a session precisely.
- Permission modes — the full detail behind Manual, Accept edits, Plan, Auto, and Bypass permissions, for tightening or loosening how much an agent can do unsupervised.
- Claude app downloads — the desktop and mobile apps.
- Groundwork agent rules — how this repo keeps one shared agent contract across tools.