NexusBerry workshop · Agentic coding

Cheat Sheet

The workflow from today's session, the rules, and every command we used.

The loop, for every feature

1. Brief. Context lives in files: CLAUDE.md for rules and commands, SPEC.md for features, "done when", and out of scope.
2. Plan. Shift+Tab to Plan mode. Read the plan, steer it, then approve.
3. Build. One feature per conversation.
4. Verify. Give Claude a check it can run (build, test, screenshot). Ask for evidence. Then test it yourself.
5. Commit. Save every green step in git.

Then /clear and start the next feature with a fresh context.

Four golden rules

• One feature per conversationSmall context, fewer mistakes.

• Plan what you can't say in one sentenceFix the approach while it's cheap.

• Never accept code you can't explainAsk "why?" until you can.

• Two failed fixes? Start fresh/clear and write a better prompt.

Install on Windows

1irm https://claude.ai/install.ps1 | iex
2claude --version
3claude

Run in PowerShell, then open a new terminal. Needs a paid Claude plan (Pro or higher). Git for Windows is recommended.

Commands and keys

A good prompt has four parts

What the feature · Where the files · Constraints what not to touch · Check how to prove it works

WEAK: make a cart

STRONG: Implement F3 from @SPEC.md. Store the cart in React Context + localStorage. Don't change the menu page. Run npm run build and show me it passes.

Words to know

• Agent

A model plus tools (read and edit files, run commands) working in a loop until the task is done.

• Context window

Everything the model can see right now. It fills up, and quality drops as it does.

• Context engineering

Choosing what goes into the context: rules files, specs, the right files.

• Hallucination

Confident but wrong output. Verification catches it.

• Spec

A written description of what to build and what "done" means.

• Permission mode

How much the agent may do without asking you.

• Checkpoint

An automatic snapshot before Claude edits, so you can rewind. It doesn't replace git.

• Vibe coding vs agentic engineering

Prompting until it seems to work, versus a deliberate loop of spec, plan, checks and commits.

Stay safe

• API keys go only in .env.local (gitignored) and are used only on the server.
• Read commands before you approve them, especially anything that deletes or pushes.
• AI wrote it, you own it. Be able to explain every change.

Official docs: code.claude.com/docs · Read the "Best practices" page first.