Source: raw/Learn_95%_Of_Claude_Code_In_30_Minutes.md (auto-caption transcript)
Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=b4Mhoe0Z4vA · Duration: 29:10 · Uploaded: 2026-09-10 · No sponsor named; prompts offered free in the description behind an optional email signup.
A single-video course on Claude Code for people who already tried it and hit limits, ignored instructions or broken code. It covers the features in the order that controls cost and quality, ends each with an “honest catch”, and closes with a seven-day learning path. Nearly everything matches first-party guidance already in this wiki, so the value is the ordering and the plan.
Key Takeaways
- Three Claudes, three jobs. The Claude app for questions; Cowork for work on your computer (sort a folder, pull invoice totals from 100 PDFs into a spreadsheet); Claude Code for writing code inside a project. “Claude Code does not suggest code. It does the work in your actual project and then shows you what it did.”
- Install with one terminal command and type
claude, or use Claude Code inside the desktop app. Point it at a folder; it can see everything inside. - Permission modes. Auto (default; a second model checks each action and stops you only for what matters), Manual (asks before every change), Accept edits, and Plan (no changes until you approve the plan). Bypass permissions never asks; he says leave it off as a beginner.
- CLAUDE.md is the biggest unlock, and the trap is stuffing it. It loads every session, so every line costs on every message and buries the rules that matter. Have Claude write it and tell it what to leave out;
/doctortrims a bloated one. It is guidance, not a cage; anything that must always happen belongs in a hook. - Watch the context meter. As the window fills, Claude forgets early instructions and re-breaks fixed things. Check with
/contextor the wheel by the chat box./compactkeeps the gist but drops detail and is not free;/clearstarts fresh. His habit: finish a task,/clear, start the next one. - Cost depends on model and context. Haiku for mechanical tasks, Sonnet as the everyday generalist, Opus for judgment, Fable for problems the others keep failing. Pick the model at the start: each model has its own cache, so switching mid-session rereads everything at full price, and Claude Code asks you to confirm. His anecdote: a client maxed out limits by using Fable for everything, including renaming files.
- Skills. Built in:
/code-reviewand/security-review. Day-one installs he names: Anthropic’s skills repo (document pack for Word, Excel, slides, PDF), Superpowers (interviews you, plans, builds test-first, sends subagents to review), and oh-my-claudecode (routes each job to a cheaper or stronger model). Skills can run shell commands, so ask Claude to read one before installing. - Subagents run in their own context and return only the answer, keeping the main session clean. They cannot talk to each other and each costs its own tokens; use them for big reading and searching jobs.
- MCP vs command line. Every connected MCP server loads instructions whether used or not; CLI tools such as
ghare lighter but less controlled. Connect MCP for tools you use often; turn off connectors unused for a month. - Six small habits. End prompts with “use ask user question until you reach clarity”; raise
/effort(low, medium, high, extra, max) or put “ultrathink” in one prompt;@a file; type the next prompt while it works and press Escape to stop; paste a screenshot or click the element; ask side questions with/btwso they stay out of history. Bonus:/batchsplits a job into up to 30 parallel pieces in separate copies of the project and opens the pull requests. “Every one of these is speed, not judgment.” - When it goes wrong. In a loop, change the instruction entirely. When it is confidently wrong, make it explain before acting. When it did too much,
/rewind. - Security Guidance plugin (Anthropic). Pattern-matches every written file for dangerous code (no model, free), diffs each turn into a background security review, and runs a deeper review on each commit using a separate, fresh-context call. It blocks nothing and can miss things.
- Hooks and Remote Control. Ask in plain English for a hook (“run my tests every time I edit a file”);
/hookslists them./remote-controlpairs the session to the Claude phone app; the work still runs on your computer, which must stay awake. - When not to use it. Tiny jobs you already know the answer to, and questions unrelated to a codebase. “Claude Code makes a mediocre engineer faster and a careless engineer dangerous.”
Try It: his seven-day path
- Install, point at a folder, make one small change in plan mode.
- Have Claude write your
CLAUDE.mdwith his “leave out what you can figure out” prompt. - Watch the context meter and
/clearbetween tasks instead of compacting forever. - Choose model and effort on purpose.
- Install the Security Guidance plugin and one skill.
- Break something on purpose and recover with
/rewind. - Give it a real job you have been putting off, turn on Remote Control and walk away.
Open Questions
- The permission-mode names come from captions (“except edit” is read here as Accept edits). Confirm in the CLI reference.
- His
CLAUDE.mdprompt was shown on screen, not read aloud.
Related
- CLAUDE.md Primer — what belongs in the file
- Context Management in Claude Code — Anthropic’s own compaction and clearing advice
- Model vs. Effort — what the two dials change
- Security Guidance Plugin — the plugin he installs on day five
- Subagents and Hooks — the two extension mechanisms
- Remote Control — steering a local session from your phone
- Claude 5.5 Mistakes and Fixes — a longer habit list from another creator
- Why Agents Burn Tokens (Sharbel) — his deeper cost video