10 stages · 81 lessons · every one version-stamped

Never opened a terminal, to shipping agents.

Not 650 lessons. 81 that are all verified against the Claude Code release they were recorded on. We lose a volume war and win a depth war — so the count is the honest one.

The spine, in one sentence

Memory is what Claude always knows. Skills are what it knows on demand. Subagents are where it runs. MCP is what it can reach. Hooks are what it cannot avoid. Everything else is scheduling, packaging, and governance on top of those five.

Stages 01 — 10

Click any stage to open it.

2 stages are marked ⚡ — material no consumer Claude community teaches, because it comes out of enterprise delivery.

  • 1.1What Claude Code actually does when you're not watchingCan explain agentic coding to a colleague in one sentenceV 7m
  • 1.2Pick your surface: Desktop, terminal, IDE, web, phoneChooses the surface that matches their skill level and installs itV 9m
  • 1.3The free-plan trap: what Claude Code actually costs⚡Knows Claude Code requires Pro/Max/Team/Enterprise/Console and is on the right oneV 5m
  • 1.4Install and verify in four minutesclaude --version returns 2.1.221 and claude doctor is cleanB 8m
  • 1.5The terminal, for people who have never opened oneNavigates directories, runs a command, doesn't panicV 11m
  • 1.6Your first real change, start to finishShips one working change to a real projectB 16m
  • 1.7Install the Claude Guys Starter KitHas the plugin installed and /claude-guys: commands workingB 6m
  • 2.1Prompting an agent is not prompting a chatbotWrites specific, verifiable prompts instead of vague requestsV 10m
  • 2.2Plan mode: Shift+Tab twice before you touch anythingPlans and approves before any code changesB 9m
  • 2.3The six permission modes, and which one you should live inPicks a mode deliberately and knows bypassPermissions is container-onlyV 12m
  • 2.4Your first allowlist: fewer prompts, same safetyShips permissions.allow/deny rules that cut prompts without losing safetyB 14m
  • 2.5Reading your context budget with /contextReads the breakdown by category and knows what to cut firstV 8m
  • 2.6/clear vs /compact: the two-second habit that halves your billClears and compacts at the right momentsV 7m
  • 2.7Undo: checkpoints, /rewind, and double-EscRecovers from a bad direction without losing the sessionB 8m
  • 2.8Model and effort: default, sonnet, opus, fable, haiku⚡Picks a model and effort level per task; knows /fast costs $10/$50 per MTok and never toggles it mid-conversationV 11m
  • 3.1/init and your first CLAUDE.mdHas a working CLAUDE.md under 200 linesB 12m
  • 3.2The four levels: enterprise, user, project, localRoutes an instruction to the correct level and knows they concatenate, not overrideV 9m
  • 3.3Splitting a fat CLAUDE.md into .claude/rules/Ships path-scoped rules that load only when a matching file is touchedB 13m
  • 3.4What Claude writes about you: auditing auto memory⚡Audits ~/.claude/projects/<project>/memory/MEMORY.md and edits itV 8m
  • 3.5CLAUDE.md vs AGENTS.md when you run more than one agent⚡Wires interop with a one-line @AGENTS.md import or a symlinkW
  • 3.6The routing decision: memory, skill, or hook?⚡Correctly routes an instruction to CLAUDE.md vs a skill vs a hookV 10m
  • 4.1Why slash commands became skills⚡Explains the merge and the three-level progressive-disclosure modelV 9m
  • 4.2Author your first skill in ten minutesShips ~/.claude/skills/summarize-changes/ and invokes it two waysB 12m
  • 4.3Frontmatter, field by fieldUses every optional field deliberately, not by cargo cultV 14m
  • 4.4Descriptions that actually fire (the 1,536-character budget)⚡Writes a description whose key use case is first, inside budgetV 10m
  • 4.5Arguments and substitutionsBuilds a parameterized skill using $ARGUMENTS, $N, ${CLAUDE_SKILL_DIR}B 11m
  • 4.6Live data with dynamic context injection !cmd⚡Ships a skill that arrives pre-loaded with git diff HEAD outputB 12m
  • 4.7Personal vs project vs plugin scopePlaces a skill at the right scope and predicts precedence conflictsV 8m
  • 4.8Multi-file skills: reference.md, scripts/, allowed-toolsBuilds a skill that runs a bundled script and returns visual outputB 16m
  • 4.9Locking down destructive skills⚡Sets disable-model-invocation: true on anything with side effectsV 9m
  • 4.10The house style guide: 500 lines, top-loaded, under budget⚡Writes to a standard that survives compactionW
  • 4.11Proving a skill works: evals with skill-creator⚡Produces a benchmark.json showing with-skill vs without-skill pass rate, tokens, and timeB 22m
  • 5.1Context is the scarce resourceExplains when a subagent beats doing it inlineV 8m
  • 5.2The built-ins: Explore, Plan, general-purposeKnows which built-in Claude reaches for and why Explore skips CLAUDE.mdV 10m
  • 5.3Write your first subagentShips ~/.claude/agents/code-reviewer.md with tool restrictionsB 12m
  • 5.4Cheap workers: tool restriction plus Haiku⚡Builds a read-only Haiku subagent for bulk classification workB 10m
  • 5.5context: fork — skills that run as subagents⚡Converts a heavy research skill into a backgrounded forked agentB 12m
  • 5.6Agent pods: fanning out without collisions⚡Runs 4 read-only subagents in parallel on one audit and merges findingsB 15m
  • 5.7The honest cost of agent teams (about 7x)⚡Decides between subagents, agent view, and agent teams on cost, not vibesV 9m
  • 6.1MCP without the word "protocol"Explains MCP to a non-technical colleague in business termsV 8m
  • 6.2Your first server: Notion in five minutesHas a working connected server verified via /mcpB 9m
  • 6.3Email, calendar, and notes: the personal stackHas three personal servers live and a working "what's on my plate" promptB 16m
  • 6.4Transports: http, stdio, sse (deprecated), wsPicks the right transport and knows SSE is deprecatedV 10m
  • 6.5Scopes and team sharing via a committed .mcp.jsonShares a server with a team through git without leaking a credentialB 11m
  • 6.6OAuth, secrets, and ${VAR} expansionAuthenticates a remote server without committing anythingB 10m
  • 6.7The MCP diet: why 25 servers makes Claude worse⚡Audits /context, disables unused servers, swaps one for a CLI toolV 13m
  • 6.8The browser and scraping stack: Playwright and FirecrawlAutomates one real web task end to endB 14m
  • 7.1Context suggests, hooks guarantee⚡Explains why CLAUDE.md is context and hooks are configurationV 9m
  • 7.2All 31 events, mapped to six cadences⚡Picks the correct event for a stated requirement, first tryV 14m
  • 7.3Anatomy: event → matcher → handler, plus if⚡Writes a correctly nested settings.json hook blockB 12m
  • 7.4JSON in, JSON out, and never both⚡Writes a hook returning permissionDecision without mixing exit code 2 with JSONB 13m
  • 7.5The guardrail: blocking rm -rf and secret reads⚡Ships a PreToolUse hook that denies a destructive command with a reasonB 12m
  • 7.6The janitor: PostToolUse formatters and linters⚡Auto-formats every file Claude edits, without askingB 10m
  • 7.7Get pinged: Notification and Stop hooks to desktop and phone⚡Gets a desktop notification and a Telegram message when Claude needs inputB 14m
  • 7.8The 18-event reference rig⚡Installs a production hook configuration covering 18 eventsB 20m + W
  • 8.1/goal: set the outcome, walk away⚡Sets a goal and watches it self-clear on completionB 10m
  • 8.2The 4,000-Character Contract⚡Writes conditions the transcript can prove, with a turn boundV 15m
  • 8.3/loop mode one: fixed intervalRuns a cron-style loop and stops it with EscB 8m
  • 8.4/loop mode two: let Claude self-pace⚡Uses self-paced mode (Claude chooses 1 min to 1 hr) where it beats a fixed intervalV 7m
  • 8.5loop.md: your always-on maintenance brief⚡Ships a project .claude/loop.md that runs unattendedB 10m
  • 8.6Choosing your scheduler: Routines vs Desktop tasks vs /loop⚡Picks correctly using the three-tier table and the 1-hour / 1-minute floorsV 12m
  • 8.7Managing scheduled tasks and the 7-day expiry⚡Lists and cancels tasks by 8-char ID; knows recurring tasks self-delete after 7 daysB 9m
  • 8.8Monitor beats polling⚡Reacts to log lines as they stream instead of burning tokens pollingB 11m
  • 8.9Worktrees: two sessions, one repo, zero collisions⚡Runs two --worktree sessions on the same repo safelyB 13m
  • 9.1Standalone or plugin?Decides correctly and understands /plugin-name:skill namespacingV 8m
  • 9.2Build a plugin from your skillsShips a plugin tested with claude --plugin-dirB 15m
  • 9.3Publish a marketplace others can add⚡Has a live GitHub marketplace installable via /plugin marketplace addB 14m
  • 9.4Get listed on anthropics/claude-plugins-community⚡Has a submission-ready plugin and knows the in-app form goes to community, not officialW
  • 9.5Output styles (and why /output-style is gone)⚡Ships a custom output style via /config, knowing the command was removed in v2.1.91B 10m
  • 9.6Statusline and settings precedence masteryHas a custom statusline and can predict which settings file winsB 12m
  • 9.7Versioning, changelogs, and not breaking your users⚡Always sets version (git distribution makes every commit a new version otherwise)W
  • 10.1Headless mode and --bare⚡Runs a reproducible scripted invocation with structured JSON outputB 12m
  • 10.2The Agent SDK in thirty lines⚡Builds a minimal Python agent using the same loop as the CLIB 18m
  • 10.3Branding rules for anything you ship to a client⚡Knows "Claude Agent" and "Powered by Claude" are permitted; "Claude Code" / "Claude Code Agent" are notW
  • 10.4CI/CD: @claude on a real repo⚡Has @claude responding on a live GitHub repoB 16m
  • 10.5Dynamic workflows: fan out and cross-check⚡Runs a script that orchestrates many subagents and cross-checks their findingsB 20m
  • 10.6Building an eval harness for a client's prompts⚡Delivers a with-skill vs without-skill pass rate, token, and latency reportB 25m
  • 10.7Cost engineering: cutting a real bill by 60%⚡Cuts measured token spend on a real project using the ten documented leversB 18m
  • 10.8Sandboxing and the security review⚡Runs /sandbox and /security-review; knows WSL1 is unsupportedB 15m
  • 10.9managed-settings.json: governance for a 200-seat rollout⚡Ships a hardened managed policy with deny rules and sandbox enforcedB 20m
  • 10.10Forward-deployed: the first 30 days inside a client⚡Runs a real engagement from audit to first three agents in productionW

Start at stage one

The first stage ends with something shipped.

Get the files

The assessment

How good are you at Claude, actually?

Eight questions. You get the level you actually test at, the one move that gets you to the next one, how many hours a year the job you name is costing you, and three skill files cut to your title, company and industry.

Take the test

Under two minutes · no account · change any answer

  • 01The Beginner
  • 02The Chat Master
  • 03The Operator
  • 04The Automator
  • 05The Agent Conductor

Section surveyed 5,013 knowledge workers: 54% rate themselves AI proficient, about 10% test as proficient.