Start here · Guide 2 of 10
Pick the right model, plan first and keep your usage under control.
In any AI tool, with Claude Code as the example
Use the smallest model that can do the job, and step up when the work is hard. Check your tool's own model page, because names and prices change.
/model switches model. The docs describe Haiku for simple tasks, Sonnet for daily coding, Opus for complex reasoning and Fable for the hardest, longest work.
/modelhaiku simple taskssonnet daily coding tasksopus complex reasoning tasksfable hardest, longest-running tasksLong conversations cost more, so start a new one for each new job. Check what you have used before starting something big.
/clear starts fresh and costs nothing. /usage shows what you have used. The next guide, Manage your usage, has the rest.
/clear# fresh start for the next job/usage# what this session and your plan have usedAgree the plan first, while changes are cheap. Ask the tool to question you, so gaps show up before anything gets made.
Shift+Tab cycles to plan mode, where it reads and explores but does not edit your source files.
Try this
Interview me about this project until you understand what I want. Ask one question at a time. Then write the plan, and do not build anything yet.
Shift+Tab cycles to plan mode# reads and explores, does not edit source filesInterview me about this project, one question at a time.If you keep typing the same instructions, save them once and reuse them.
A skill is a SKILL.md file in .claude/skills/<name>/, which you run with /name. Put it in ~/.claude/skills/<name>/ instead to use it in every project. It only loads when used, so long instructions cost almost nothing until then.
.claude/skills/review/SKILL.md# this project only~/.claude/skills/review/SKILL.md# every project/reviewMost AI tools have somewhere for standing instructions. Put your rules and project details there, and keep them short.
It reads CLAUDE.md at the start of every session, and /init creates a first draft. The docs advise under 200 lines, and warn it is context rather than enforced rules. Rules for every project, such as which skill to use for which job, go in ~/.claude/CLAUDE.md.
~/.claude/CLAUDE.md# your rules, every project./CLAUDE.md# this project, shared with your teamSee @docs/setup.md for how to run it.# an @ line pulls in another fileInstalling a skill does not mean it gets used. Check that each one loads where you need it, and say when to use it.
Skills in one project's .claude/skills are not visible in another. A skill with disable-model-invocation: true only runs when you type its name. /context shows what is filling the conversation, skills included.
/context# what fills the conversation, skills includeddisable-model-invocation: true# this skill runs only when you type its nameHand a job to an agent when it is noisy, can run in parallel or needs fresh eyes. Keep simple work in your main conversation, because extra agents cost extra usage.
A subagent has its own context and counts toward the same limits. The docs say that by default you get parallelism or a fresh reviewer, but not a smarter answer.
main conversation├─ subagent own context, reviews the diff└─ subagent own context, searches the docs# both count toward the same usage limitsStart with approvals on, and allow more only for things you trust. Anything a tool can reach, a mistake can reach too.
It enforces its permission rules itself, in the order deny, ask, allow. The docs say to use bypass mode only in isolated environments such as containers or VMs.
{ "permissions": { "deny": ["Read(./.env)"], "ask": ["Bash(git push:*)"], "allow": ["Bash(npm run test:*)"] }}# checked in the order deny, ask, allowConnecting your calendar, files or database makes an AI tool far more useful, and widens what can go wrong. Connect only what you need, from people you trust.
It uses MCP for this. The docs warn that servers fetching outside content can expose you to prompt injection.
claude mcp add --transport http notion https://mcp.notion.com/mcp# connect only servers you trustRead it as a colleague would. Anthropic's test is that if they would be confused, the model will be too.
Write down what you are working on and how you like things done.
Ask for a plan, correct it, then let the tool start.
Offer something it can run that passes or fails, and ask for evidence such as the output or a screenshot.
A clean conversation stops the next task paying for the last one's history.
A standing instruction is context the model reads, and it can be ignored. Use real permission settings for anything that must never happen.
Each agent sends its own requests, so a big fan-out can use up a weekly limit quickly.
Model names, limits and commands move often. Everything here was checked on 1 November 2026, so confirm it on the linked page.
Notes, evidence and sources
Musa Burke's own notes from building this guide and other sites with Claude Code, not study findings.
My first two designs came out badly because the agents building them used none of the design skills I had set up. The brief never told them to.
I gave Claude a weekly usage limit and pasted in my /usage at checkpoints, because the session could not read it. One research agent alone used about 261,000 tokens.
I rejected several versions of this site after every automated check had passed. Only looking at it told me it was boring.
This guide exists because Claude finally asked me questions before building. My answers changed what the site was for.
1. Claude Code docs: models · Checked 1 November 2026
2. Claude Code docs: costs · Checked 1 November 2026
3. Claude Code docs: costs and /usage · Checked 9 November 2026
4. Claude Code docs: plan mode · Checked 1 November 2026
Claude Code permissions page · costs page · best practices page
5. Claude Code docs: skills · Checked 1 November 2026
6. Claude Code docs: CLAUDE.md · Checked 1 November 2026
7. Claude Code docs: CLAUDE.md imports · Checked 9 November 2026
8. Claude Code docs: subagents · Checked 1 November 2026
9. Claude Code docs: permissions · Checked 1 November 2026
10. Claude Code docs: MCP · Checked 1 November 2026
11. Anthropic prompting guide · Checked 1 November 2026
12. Claude Code docs: best practices · Checked 1 November 2026