R140 · AI education › Teaching materials and glossaries

Claude Code in 10 levels: the technical steps, checked for teaching

Read the original

teachingharness-configagent-harnessagent-skillscontext-engineeringlearning-resources

Source and verdict

A 17-minute (1,033 s) beginner walkthrough that builds a full Claude Code setup on one empty repo in ten levels, each adding one mechanism: config, memory, prompting, project hygiene, tools, outputs, saved process, delegation, habits, automation. It is good teaching material because the order is a dependency order (each level needs the one before) and every level ends in something a learner can run. It is a feature tour, not a design guide: it says what each piece is and roughly when to reach for it, but little on failure modes, cost or governance.

The ten levels as technical steps

Each level is what you type or create, in the video's order. Her closing mnemonic: install, remember, ask, organize, connect, show, repeat, delegate, hand off, schedule. She says Levels 1 to 4 take about ten minutes and give the biggest jump.

  1. Setup: install, scaffold, configure.

    • Install with the one-line installer (curl -fsSL https://claude.ai/install.sh | bash), check with claude --version, then run claude doctor.

    • Never start from an empty folder: scaffold new repos from a Copier template that already holds CLAUDE.md, a .claude/ folder with one agent, one skill and one command, and .mcp.json. copier update later pulls template improvements into existing repos.

    • Three instruction scopes: ./CLAUDE.md (committed, team rules), ~/.claude/CLAUDE.md (your rules for every project), ./CLAUDE.local.md (yours, this checkout, gitignored).

    • Two settings files: .claude/settings.json (permissions, hooks, env) and .claude/settings.local.json (personal overrides). Rule of thumb: markdown says how to work, settings say what is allowed. Keep CLAUDE.md short.

  2. Memory: import, then separate rules from facts from references.

    • Import what another assistant knows about you (Claude settings, memory import: Claude gives a prompt, you run it in ChatGPT, paste the result back; two to three minutes).

    • From another coding tool, merge .cursorrules or AGENTS.md into CLAUDE.md, or pull it in with an @path import.

    • Auto memory saves learned facts (such as how you run tests) unasked; /memory shows and edits them.

    • Layout: CLAUDE.md = rules you set; auto memory = facts Claude learned; docs/ = reference material. Add claude-mem (session recall) or a graph memory such as Graphiti MCP only when you see something being forgotten.

  3. Prompting: make it ask first. Add "Before you answer, ask me what you need." to the prompt, so a wrong question gets caught before work starts.

  4. Projects: isolate, scope, and manage the context window.

    • One git worktree per parallel feature, each on its own branch with its own session.

    • CLAUDE.md sections: stack (language, tools, cloud), build, test and lint commands (e.g. pytest, ruff), project structure, folders not to touch, definition of done.

    • A nested CLAUDE.md in any folder with special rules; it loads when Claude works there and keeps the root file small.

    • Commit .claude/agents, .claude/skills, .claude/commands and .mcp.json so every clone gets the same setup.

    • Name sessions so /resume finds them.

    • When answers degrade over an hour, the window is full (system prompt, CLAUDE.md, tool definitions, every file read, the conversation). Inspect with /context; then /compact (summarize), /clear (fresh start) or /rewind (code and conversation back to a checkpoint).

    • Pick the model at the start: a mid-session switch rebuilds the prompt cache. The Context Mode plugin keeps big raw outputs (logs, browser snapshots) out of the window.

  5. Tools: MCP, connectors, CLIs, browser.

    • claude mcp add with a scope: local (you, this project), project (written to .mcp.json, shared), user (all your projects). Prune unused servers. Build an internal one with Anthropic's mcp-builder skill.

    • Connectors (Gmail, Calendar, Drive) for work data. Dev servers she rates: Context7 (current library docs), GitHub (issues, PRs), Apify (scraping); in companies Slack, Jira, Confluence. Supabase and Stripe publish official skills for auth, databases and payments.

    • Prefer a CLI when a good one exists (gh), and add trusted commands to the permission allow list in settings.json so they run without prompts. She cites one comparison where a CLI used 35 times fewer tokens than MCP for the same task.

    • Browser: Claude in Chrome for pages with no API (an internal dashboard); Playwright for repeatable app tests.

  6. Outputs: something you can see. Ask for an artifact in the Claude app, or one HTML report in Claude Code (a test summary, an architecture diagram). Add a custom status line showing model, context fill and session cost (/statusline). Mods add small panels, e.g. one that previews what a command would delete.

  7. Your process: skills, hooks, plugins, workflows.

    • A skill is a folder under .claude/skills/ with a SKILL.md. Her starter set: skill-creator, Superpowers (plan first, test first), Matt Pocock's grill-me and handoff, GSD (state for long projects), Context Mode, frontend-design, Humanizer. Built-in reviews: /review, and a deeper cloud review with ultrareview. Install two or three; each costs context. Vet third-party skills first.

    • Hooks give the guarantee CLAUDE.md can't ("always run the linter" is followed most of the time). Defined in settings JSON: before a tool runs (block dangerous commands, protect .env), after each edit (formatter, linter), on stop (notify you), on session start (load fresh context).

    • A plugin bundles skills, agents, commands, hooks and MCP servers: /plugin marketplace add, then install. Teams keep plugins in a private marketplace repo.

    • A workflow runs several steps as one (gather, draft, check, write); saved in .claude/workflows/, watched with /workflows. Deep research is a built-in example.

  8. Scale: goals, headless runs, subagents.

    • /goal <condition> keeps working until the condition holds. Make it checkable: "page loads in under one second", not "make the page faster".

    • claude -p runs one task from a script or CI, e.g. a GitHub Action that reviews every PR. The Agent SDK runs the same loop in your own code.

    • A subagent is a markdown file in .claude/agents/ (frontmatter: name, when to use, allowed tools) with its own context; it returns a short result. Her template ships a read-only code reviewer. Built-ins include Explore and Plan. Send messy side tasks to subagents; give parallel code-writing agents their own worktrees.

    • Agent teams (experimental): teammates message and debate rather than only report back.

  9. Habits. /insights (a report on your last 30 days of use), /btw (side question without interrupting the running task), /voice (dictation). With the status line and a stop-notification hook you can stop watching the screen.

  10. Automation: run without you.

    • /loop repeats a prompt on an interval in the open session.

    • Routines (now called cloud scheduled tasks) run a Claude Code prompt in the cloud on a schedule, like cron for an agent; Cowork scheduled tasks do the same for non-code work. Example: every Friday, review the week's changes and draft release notes.

    • Dispatch: send a task from your phone that runs on your computer.

    • Remote Control: drive a session already running on your computer from phone or web.

    • Channels: push Telegram or Discord messages into a running session (claude --channels ..., pair the bot); research preview, so lock it to an allowlist.

The finale chains them: from her phone she triggers the Friday routine; the agent reads the project rules, plans the change, hands review to a subagent and posts a summary to the team channel.

Corrections and caveats

Most of the video matches the current Claude Code docs as of 2026-11-18. Eight points are out of date, overstated or missing a condition a learner will trip on.

Video saysCurrent docs saySource
Every connected MCP server costs tokens on every messageMCP tool definitions are deferred by default: only tool names and server instructions load until a tool is used. CLIs are still leaner (no per-tool listing), so the advice holds, but the penalty is smaller than impliedCosts
A CLI used 35x fewer tokens than MCPUnsourced in the video, and likely measured before tool deferral. Treat as directionalnone
Create subagents with /agent/agents now only prints a reminder (v2.1.198 and later); ask Claude to create one, or edit .claude/agents/ directlyCommands
/ultra review/ultrareview [PR or branch], or /code-review ultra; /review is an alias of /code-review with effort levelsCommands, ultrareview
RoutinesRenamed cloud scheduled tasks. Three schedulers exist: cloud scheduled tasks (laptop can be closed), Desktop scheduled tasks (your machine), /loop (only while the session is open)Scheduled tasks
/goal runs until the condition is metA separate evaluator judges the condition from the transcript only; it does not run commands or read files. "Loads in under one second" works only if Claude measures and prints it. Bound it with a clause such as "or stop after 20 turns"Goal
CLAUDE.local.md for your own rules, plus one worktree per featureA gitignored CLAUDE.local.md exists only in the worktree where you made it. For personal rules across worktrees, import a file from your home directoryMemory
Agent teams and channels just workAgent teams are disabled by default (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1). Channels are a research preview (Telegram, Discord, iMessage), and Team and Enterprise orgs must enable themAgent teams, Channels

Smaller notes. Target under 200 lines per CLAUDE.md; path-scoped rules in .claude/rules/ are the docs' answer to "one part of the repo has special rules", alongside nested CLAUDE.md files. /insights reads local history, so it is not available in cloud sessions. The built-in alternative to a Copier template is /init (CLAUDE_CODE_NEW_INIT=1 adds skills, hooks and personal memory). The model-switch cache warning is correct, and /usage now shows a prompt-cache line that names likely miss causes.

What's missing for a staff-level audience

The video adds capability at every level and never removes or measures any. A staff engineer needs the other half:

How to teach it

Use the video as pre-work or the first 17 minutes, then make learners build the ladder on one shared toy repo. Each block ends in an artifact you can check. Two sessions of about 90 minutes work better than one long one.

Session 1: Levels 1 to 5 (the setup every engineer should have).

  1. (15 min) Install, claude doctor, /init. Write a CLAUDE.md under 30 lines with stack, commands, do-not-touch folders and definition of done. Pair review: cut anything Claude could learn from the code.

  2. (15 min) Memory triage: give learners ten mixed statements and have them sort each into CLAUDE.md, auto memory, docs/, or settings. Then show /memory.

  3. (10 min) Ask-first prompt: run the same vague request with and without "ask me what you need first" and compare.

  4. (20 min) Context lab: read a large file, run /context, then /compact, /clear and /rewind. Switch models mid-session and read the cache line in /usage.

  5. (20 min) MCP vs CLI: do one GitHub task through the GitHub MCP server and once through gh, and compare /context and tokens. Add the gh commands to the allow list.

  6. (10 min) Debrief on the corrections table above.

Session 2: Levels 6 to 10 (making it repeatable and unattended).

  1. (20 min) Hooks over rules: write "always run the linter" in CLAUDE.md, then the same as a PostToolUse hook. Add a PreToolUse hook that blocks edits to .env, and prove it fires.

  2. (20 min) One skill, measured: write a small skill with skill-creator, run three tasks with and without it, keep it only if it helps.

  3. (20 min) Delegation: a read-only reviewer subagent; then /goal with a condition Claude must print evidence for and a turn cap. Discuss what the evaluator can and cannot see.

  4. (20 min) Unattended: a claude -p script with a deny list and a turn cap, then the same prompt as a cloud scheduled task. Ask: what could it do that you would not want, and which setting stops it?

  5. (10 min) Debrief on the missing-for-staff list.

Discussion questions. Which levels give guarantees (settings, hooks, deny rules) and which only make behaviour likely (CLAUDE.md, skills, prompts)? When does a CLI beat MCP, and when is MCP worth its cost? What is the smallest setup that still gets the Level 1 to 4 gains?

Sources

Connections

See R140 on the knowledge map

Part of the Research notes on nowucca.com. Figures and claims are the source's own unless the summary says otherwise.