Claude Code tips for beginners

The habits that matter most: restart Claude Code after installing skills or MCP servers, start a fresh session for each big step, press Esc before sending a normal message when the slash menu is open, and let subagents do long research in their own context. These come from building sites with Agent2Ship; each one fixes a mistake that is easy to make.

1. Restart after every install

After installing skills or MCP servers, close Claude Code completely and open a new session. MCP servers and newly added tools load reliably only in a new session, and building in the same session as installing is the main reason kit-built sites look basic.

2. Use a fresh session for each big step

Build, design and deploy each work best in a new session. You don't lose anything: Agent2Ship writes every decision into .claude/agent2ship/ (brief, stack, design plan) and every change into seo-reports/changelog.md, so the next session reads the files instead of a long chat history.

If the / command list is open and you press Enter, your message becomes the highlighted command. Press Esc first, then send your message.

Claude Code tips for saving tokens

Let cheaper models do research. In Agent2Ship, 10 of the 15 subagents (research, planning, SEO checks, launch, growth) are set to Sonnet in their files, so they use less of your plan than your main model would.
Keep sessions short and focused. A fresh session with good files costs less than a long one carrying old context.
Ask for edits in place. Asking Claude to change a file directly is cheaper than having it paste whole files into the chat.
If you hit a usage limit, wait for the reset and run the same command again. The design step resumes from DESIGN-PROGRESS.md instead of starting over.

Claude Code subagents best practices

One job per subagent. A keyword researcher, a schema specialist, a fact-checker: narrow jobs give better, checkable output.
Use them for long, noisy work. A subagent works in its own context window, so a crawl or a competitor study doesn't fill your main conversation.
Pick the model per agent. Research and checking can run on Sonnet; building and writing can use your session's model.
Make them write results to files. Reports in seo-reports/ and decisions in .claude/agent2ship/ survive restarts and can be reviewed.
Don't let any agent approve for you. Publishing pages, deploying, or changing URLs, redirects, robots.txt or noindex should always wait for your own yes.

More Claude Code tips and tricks

Check Python first. The UI UX Pro Max design skill needs Python 3. Run python --version before /a2s-build.
Approve the Astro docs MCP once. It's added at project scope (.mcp.json); after the restart, Claude Code asks you to approve it.
Allow Playwright for design reviews. About 150 MB. It lets the kit take desktop and phone screenshots. Without it, you check the site yourself.
Test deploys are noindex. testDeploy: true keeps Google away from your test URL until the real domain is live.
Test platform facts yourself. When the kit asks you to check something on a platform (a limit, how emojis count), do it, and say whether you tested on the web or the app.
Ship pages in small batches. 3–5 pages, then measure. Mass-producing near-identical pages is what Google's spam policies target.

Stuck on an error? See troubleshooting. New here? Start with getting started.