How To Use Claude Code To Build Anything (for beginners)
Decision Card
Effort: One evening (~1–2 hours) — install the Claude Code desktop app, point it at an empty folder, build the three-panel Next.js dashboard on dummy data in plan mode, then wire one panel to Google Calendar through the Zapier SDK.
Honest take: The video is sponsored by Zapier, and the “easy backend” section is the sponsor’s product; it never mentions that routing your Gmail/Calendar through Zapier adds a third-party account, a dependency and a possible cost. The blanket advice to “always choose auto and skip the permissions” overstates what Anthropic claims: its docs call auto mode a research preview that reduces prompts but does not guarantee safety.
Concrete next steps:
- Read the official permission-modes page before copying the “always auto” habit (10 min): https://code.claude.com/docs/en/permission-modes
- Reproduce the dummy-data-first workflow: plan mode → answer the clarifying questions → accept → check in the browser pane (~30 min, most of it generation time).
- If you want live data, follow the Zapier SDK quickstart (Node.js 20+,
npm install @zapier/zapier-sdk, browser login) and swap one panel only (~30 min): https://docs.zapier.com/sdk/quickstart - Skip if you already use Claude Code with plan mode and git; the only new material is the Zapier SDK walkthrough, which the quickstart covers faster.
TL;DR
A beginner walkthrough of the Claude Code desktop app: pick a folder, set the permission mode, model and effort, plan and build a Next.js “morning summary” dashboard on dummy data, then replace one panel with real Google Calendar events via the sponsor’s Zapier SDK. The workflow is sound, but the transcript is a rough machine translation (“Cloud Code” for Claude Code) and several claims are loose or sponsor-driven.
Key Points
- Every project is just a folder on disk; you pick or create one before starting, and all generated code lands there. 00:38
- The app offers local, cloud, remote control and SSH environments; the video uses local only. 01:20
- Permission modes set how much Claude may do without asking; the presenter always picks auto. 01:43
- Reasoning effort trades speed and usage against quality; “high” is presented as enough for most iterative coding. 02:16
- Model advice: skip Haiku for coding, use Opus 5 as the middle ground, and switch to Fable 5 at max effort when stuck. 03:26
- Build the UI on realistic dummy data first so you can judge layout and behaviour before touching real integrations. 04:53
- Use plan mode for anything new; its question-and-answer step lets you correct the design before code is written. 06:43
- The first build took about 7 minutes; the result shows in the in-app browser pane or in any browser via the localhost URL. 07:22
- The Zapier SDK handles OAuth for you: paste its docs into the chat, ask to replace one panel with real data, and approve the browser login. 09:20
- Use version control so you can roll back bad AI changes, and open a second chat to run a parallel task while the first one builds. 11:27
Notable Quotes
“personally I usually always choose “auto” and skip the permissions because the models are so good now that it’s extremely reliable.” 02:00
“You don’t need to step on the gas if you’re just changing the color of your landing page.” 04:03
“when you are building something from scratch, developing new features or a new concept, use planning mode.” 06:43
Verified Claims
1. Auto mode lets Claude act without asking, and is “extremely reliable”.
- 01:56
- Claude Code permission modes
- Verdict: Disputed in part — auto mode exists and a classifier reviews each tool call instead of prompting you, but the docs describe it as a research preview that does not guarantee safety.
2. Plan mode makes Claude plan (and ask questions) before writing code.
- 05:52
- Claude Code permission modes
- Verdict: Confirmed — Claude explores and writes a plan without editing source; approving the plan switches to the mode you choose.
3. The desktop app offers local, cloud and SSH environments plus a built-in browser preview.
- 01:20
- Claude Code desktop docs
- Verdict: Confirmed — the docs list Local, Cloud, SSH (and WSL on Windows) and a Browser pane that runs your dev server. The pane’s tools are not available in SSH sessions.
4. You can run parallel chats/agents on separate tasks.
- 11:54
- Redesigning Claude Code on desktop for parallel agents
- Verdict: Confirmed — the redesigned app has a sidebar for multiple sessions, each able to run in its own git worktree.
5. Lower effort is faster but lower quality; higher effort costs more time and usage.
- 02:20
- Claude Code effort levels (third-party)
- Verdict: Inconclusive — the trade-off matches third-party guides, but I did not retrieve the official effort documentation, and those guides disagree on defaults. “High is enough” is the presenter’s own untested-here rule.
6. Haiku is not worth using for code (“two-year-old level”).
- 03:15
- Claude Haiku (Anthropic)
- Verdict: Disputed — Anthropic positions Haiku as a fast, cost-efficient coding sub-agent, and said Haiku 4.5 gave similar coding performance to Sonnet 4 at a third of the cost.
7. Next.js is “just a programming language”.
- 04:36
- Next.js installation guide
- Verdict: Disputed — Next.js is a framework scaffolded with
create-next-app, not a language.
8. Zapier has over 9,000 integrations.
- 08:50
- Zapier SDK quickstart, Zapier developer page
- Verdict: Confirmed, by Zapier’s own figures only (“9,000+ apps”); I found no independent count.
9. The Zapier SDK takes care of OAuth so you don’t write it yourself.
- 09:14
- Zapier SDK quickstart, Zapier SDK docs
- Verdict: Confirmed — setup is
npm install @zapier/zapier-sdkplus a browser login, and the platform handles token refresh. Server deployments need separate client credentials, which the video skips.
10. GitHub is free software that provides version control.
- 11:28
- Git vs. GitHub (Tower)
- Verdict: Disputed — Git is the version control system and works offline on its own; GitHub is a hosting service built on top of it.
Tools, Papers & Standards Mentioned
- Claude Code desktop app — https://code.claude.com/docs/en/desktop
- Claude Code permission modes (auto, plan) — https://code.claude.com/docs/en/permission-modes
- Claude Haiku — https://anthropic.com/claude/haiku
- Claude Sonnet, Opus and Fable — named in the video; no official page retrieved in this research
- Next.js /
create-next-app— https://nextjs.org/docs/app/getting-started/installation and https://nextjs.org/docs/app/api-reference/cli/create-next-app - Zapier SDK (
@zapier/zapier-sdk) — https://docs.zapier.com/sdk - Git / GitHub — https://www.git-tower.com/learn/git/faq/git-vs-github (no official page retrieved)
- OAuth, Gmail, Google Calendar, Jira, HubSpot, Slack, end-to-end testing — mentioned; no canonical source retrieved
Follow-up Questions
- What does the Zapier SDK cost per action at personal-dashboard volumes, and how does it compare with calling the Google Calendar and Gmail APIs directly?
- What does auto mode’s classifier actually block, and how often does it let through a destructive action in practice?
- How large is the measured quality gap between effort levels and between Opus and Fable on small app-building tasks, relative to the usage cost?
Sources
- https://code.claude.com/docs/en/permission-modes
- https://code.claude.com/docs/en/desktop
- https://claude.com/resources/articles/claude-code-desktop-redesign
- https://www.dsebastien.net/claude-code-effort-levels/
- https://anthropic.com/claude/haiku
- https://nextjs.org/docs/app/getting-started/installation
- https://nextjs.org/docs/app/api-reference/cli/create-next-app
- https://docs.zapier.com/sdk/quickstart
- https://docs.zapier.com/sdk
- https://zapier.com/z/Kx/developer
- https://www.git-tower.com/learn/git/faq/git-vs-github