Claude Code From Zero — 12-Step Roadmap
Install Claude Code, configure permissions, run your first plan-mode task, and ship a small PR — a zero-to-shipping roadmap for operators new to agentic coding.
Who this roadmap is for
You write code sometimes — or manage people who do — but Claude Code feels like a black box. These twelve steps take you from install to one merged change you understand. Budget one focused afternoon.
Stack assumed: macOS, Linux, or Windows with Git and Node installed. Claude Pro, Max, Team, or API access required.
The 12 steps
Step 1 — Install the CLI
Follow official install at code.claude.com for your OS. Verify:
claude --version
If command not found, fix PATH before continuing — everything else fails silently.
Step 2 — Authenticate
Run claude and complete browser login. For teams, use the org SSO your admin configured.
Step 3 — Open the right folder
cd into a real repo — not your home directory. Claude needs project context.
cd ~/projects/my-app
claude
Step 4 — Run /doctor
Inside Claude Code:
/doctor
Fix flagged issues (Node version, git, permissions). /checkup is the same command.
Step 5 — Read the permission modes
Learn the cycle: Manual → Accept Edits → Plan → Auto (names may vary by version). Start in Plan or manual for week one.
Shift+Tab often cycles modes — check /help on your build.
Step 6 — Create CLAUDE.md (project brief)
In repo root, add short instructions:
# Project context
- Stack: Next.js 15, TypeScript, Tailwind
- Test: npm run test
- Lint: npm run lint
- Never commit .env files
- Prefer small PRs; explain plan before multi-file edits
Claude reads this every session. Keep it under 80 lines.
Step 7 — First session — read only
Read the repo structure. Summarize: entry points, routing, and test command.
Do not edit any files.
Confirm it found the same architecture you expect.
Step 8 — Plan mode task
Pick a tiny issue — typo, copy change, or small UI tweak.
Plan only — do not edit yet.
Task: [describe change]
List files you will touch and how you will verify.
Wait for approval.
Review the plan like a tech lead. Say yes or refine.
Step 9 — Execute with review
After approval, allow edits. Watch diffs. Run tests yourself:
npm run lint && npm run test
Do not merge what you have not run locally.
Step 10 — Git discipline
Either let Claude commit with your message template or commit manually:
git checkout -b fix/copy-tweak
git add -p
git commit -m "fix: clarify CTA copy on pricing"
Agents excel at branches; you own the merge decision.
Step 11 — Optional — Desktop app
Install Claude Desktop → Code tab. Import a CLI session with /desktop when you want visual diffs, terminal panes, and parallel sessions. Same config as CLI.
Step 12 — Ship and retrospective
Open PR, note:
- What the agent did well
- Where you corrected it
- What to add to
CLAUDE.mdfor next time
One shipped PR beats ten playground chats.
Common blockers
| Problem | Fix |
|---|---|
| Agent edits wrong files | Narrow scope in prompt; name paths |
| Infinite tool loop | Stop session; replan with smaller steps |
| Tests fail | Paste failure output; ask for fix, not rewrite |
| Permission fatigue | Use Accept Edits only after plan approved |
Copy-paste starter prompt
You are working in this repo. Read CLAUDE.md first.
Task: [one sentence]
Constraints: small diff, run npm run lint after edits
Process: plan → wait for OK → implement → summarize changes
Graduation checklist
-
/doctorclean -
CLAUDE.mdcommitted - One plan-mode → execute cycle completed
- Tests run locally before push
- You can explain every line in the diff
Next guides
- Claude Desktop Cheatsheet
- Boris Cherny — Practical Principles
- GitNexus Guide — give agents codebase graph context
The Gear Drop and the free library carry the next milestones — MCP setups, review loops, and agent ops for MyGearHut readers.
Need this architecture deployed in your organization?
MyGearHut consults and builds custom AI agents, automated operations pipelines, and private inference infrastructure.