/level — the 10 levels of Claude Code

Situation. Most people using Claude Code have no idea how much of it they're actually using — they type prompts into a terminal and stop there, or they've built a deep system and can't see the gaps. Complication. There's no honest, repeatable way to answer "how good is my setup, really, and what's the single next thing to build?" Generic "10 levels of AI" lists don't measure Claude Code specifically, and self-assessment drifts. Question. Where are you on the Claude Code mastery curve, and what's the one move that levels you up? Answer. This skill scores you 0-10 on a ladder calibrated to the 4 systems pillars, benchmarked against the most advanced documented setup + the latest Anthropic features, then hands you a focused roadmap and a shareable dashboard.

This is NOT the generic "10 levels of Claude" (browser, desktop, projects). It's specifically Claude Code mastery. Level 0 is "just installed it." Level 10 is genuinely rare — it's the documented ceiling in the premium reference.

Pick a front door → Score it → Show the roadmap → Render dashboard → Build first step
       ↓ ↓ ↓ ↓ ↓
   quiz or scan 0-10 + 4 pillars next level only HTML (shareable) optional

Designed to be re-run. Every time you level up, the score updates.

---

Modes and flags

Parse the user's input for flags. Defaults shown.

Flag	Effect
--mode beginner (default)	"Knowledgeable friend" voice (see tone rules below). Default brand = GTM Engineer School (or template defaults). Default front door = quiz.
--mode internal	Operator-terse voice for self/team audits. Default brand = Genesys Growth kit. Default front door = scan.
--brand-kit <path>	Brand the HTML dashboard from a specific brand kit (a DESIGN.md with token frontmatter). Overrides the mode's default brand. --brand-kit none forces the plain dark defaults.
--format html (default) | ascii | both	Output shape. html = the branded dashboard. ascii = a fixed-width, paste-anywhere result card (best for sharing — Slack / LinkedIn / a code block). both = render each.
--quiz	Force the interactive quiz front door (works with zero setup; shareable).
--assess	Force the silent file scan, then stop after the assessment (no roadmap/build).
--build	Skip assessment; jump straight to building the next step (asks current level first).

If no flag: use --mode beginner and offer both front doors ("Want the quick quiz, or should I scan your setup?"). When you have a brand kit (via --brand-kit or the mode default), offer to apply it: "Brand the dashboard with your {brand} kit, or keep the plain dark theme?"

---

Phase 1 — pick a front door, then score

Both doors converge on the same result: a level (0-10) + a 0-4 sub-score per pillar. Read the premium reference for the ladder, the 4-pillar mapping, and the scoring rules ("highest level meeting ALL criteria"). It is the source of truth for both doors.

Phase 1a — Quiz (default in beginner mode)

Read the premium reference for the question bank + the answer→(level band + pillar) scoring map.

1. Ask the ~8 questions using AskUserQuestion (one tool call, multiple questions, clean clickable options). Keep it fast and fun.
2. Map each answer to its pillar + level band per the scoring map.
3. Reconcile to a single level + 4 pillar sub-scores per the rule in quiz.md (a pillar's sub-score is gated by the highest rung that pillar's answers clear; the overall level is the highest rung where ALL its required pillars clear).
4. The quiz needs nothing on disk — it works for a total beginner or a prospect.

Phase 1b — Scan (default in internal mode)

Silently inspect the environment — do NOT ask permission per check, just read what's available. Use the scan checklist in the premium reference. Cover: CLAUDE.md (all locations + depth), .claude/skills + .claude/commands (count + complexity), MCP config (.mcp.json + user config + plugins — count + categories), memory/rules/ontology depth, hooks + agents (quality-gate + orchestration signals), headless/SDK + plugin signals, browser MCPs, worktrees + /workflows, scheduled/background agents.

Then ask the 3 context questions:

1. What do you mainly use Claude Code for?
2. What's your biggest friction right now?
3. If you could automate one repeated thing, what would it be?

Determine the level via the calibrated scoring table in levels.md. Quantity ≠ level — 50 trivial skills is still Level 3; integration + complexity is what moves the rungs.

Converge — present the assessment

Anchor the ceiling using the premium reference: Level 10 = the documented most-advanced setup + every current Anthropic-native feature. Present, in the mode's voice:

• The level number + its fun name + one-line vibe.
• The dual lens: a quick per-pillar read (which of Context / Skills / Integrations / Orchestration is strong, which is the gap).
• Walk the rungs they've cleared bottom-to-top, naming the specific thing found for each, then name the first rung that stops them and why it matters (plain language, not jargon).
• "Where that puts you" — most land 1-4; 5 is strong; 7+ is rare.

Tone rules (beginner mode): talk like a knowledgeable friend, not a grading rubric. Explain why each thing matters, not just that it was found. Say so genuinely when something's impressive; say it constructively when something's messy. Define jargon inline ("headless mode — running Claude from a script instead of typing into the terminal"). Use "you/your".

Tone rules (internal mode): operator-terse. Lead with the weakest pillar and the single highest-leverage fix. Skip the encouragement.

If --assess, stop here. Otherwise continue.

---

Phase 2 — the roadmap (next level only)

Show ONLY the next-level transition — focus is everything, don't dump all 10. Read the matching section in the premium reference for the assessed level and present its 3-4 steps with time estimates, recalibrated to current Anthropic-native features (plugins, /workflows, subagents, hooks, scheduled agents, the Agent SDK, output styles, worktrees, the memory tool).

---

Generate the dashboard

Always produce output after Phase 1 (+ Phase 2 unless --assess). Pick the shape from --format (default html). For the quiz / lead-gen path, offer the ASCII card — it's the paste-anywhere artifact (Slack, LinkedIn, a code block).

HTML dashboard (--format html, default — or both)

1. Read the premium reference.

2. Resolve the brand. Pick the brand kit (a DESIGN.md with token frontmatter):

   • --brand-kit <path> given → read that file. --brand-kit none → skip to the template's dark defaults.
   • else --mode internal → projects/genesys/brand/0626-brand-kit.md (the latest *-brand-kit.md under projects/genesys/brand/).
   • else --mode beginner → a GTM Engineer School kit if one exists, else the template's dark defaults.
   • For a client run, prefer projects/consulting/active/{client}/brand/*-brand-kit.md.

3. Map brand-kit tokens → dashboard tokens (per the template's brand-token comment):

   Dashboard	DESIGN.md source	Notes
   {{ACCENT}}	colors.accent	primary (number, current dot, bars, links)
   {{DONE}}	colors.tertiary or colors.success	cleared dots/rows
   {{BG}} / {{SURFACE}} / {{SURFACE_2}}	colors.background / surface / surface-elevated
   {{BORDER}}	colors.border
   {{TEXT}} / {{TEXT_MUTED}} / {{TEXT_SUBTLE}}	colors.on-background / -muted / -subtle
   {{FONT}} + {{FONT_LINK}}	typography.*.fontFamily	add the font's web <link> (e.g. Fontshare for General Sans); else ""
   {{BRAND_NAME}} / {{BRAND_URL}}	name / domain

   If the kit is a light theme, keep text/background readable (the template assumes a dark canvas — for light brands, swap {{BG}}→light, {{TEXT}}→dark and sanity-check contrast). Honor design-production.md token-cite discipline: use the kit's exact hex, never approximate.

4. Replace every {{TOKEN}} with real assessment data — per-pillar bars + the shareable result card (level + name + vibe, sized to screenshot).

5. Write to ~/Desktop/claude-code-level.html.

6. Open it platform-aware (open on macOS, xdg-open on Linux, start on Windows/WSL).

7. Tell the user it's open + print the path + name the brand applied.

ASCII card (--format ascii — or both)

A fixed-width, monochrome, copy-paste card — the most shareable shape. Build a JSON payload and pipe it to the generator (it asserts every line is identical width, so the box never misaligns):

echo '{"level":N,"name":"<level name>","vibe":"<vibe line>",
"pillars":{"Context":c,"Skills":s,"Integrations":i,"Orchestration":o},
"takeaway":["short line 1 (≤48 chars)","line 2","line 3"],
"brand_name":"<from brand kit name>","brand_url":"<from brand kit domain>"}' \
  | python3 the premium reference

It prints the card and saves ~/Desktop/claude-code-level.txt. Then also paste the card inline in a fenced code block so the user can copy it straight from chat. brand_name/brand_url come from the resolved brand kit's name + domain (ASCII is monochrome — brand presence is the wordmark, not colour). Keep takeaway to 2-3 lines ≤48 chars each.

---

Phase 3 — build it now (optional)

Ask: "Want me to build the first step of your roadmap right now?" If yes, execute the matching build step in the premium reference for the assessed level (create the CLAUDE.md, wire the first MCP, write the first skill, scaffold memory, upgrade a skill to multi-phase, write the first headless script, etc.). Adapt to the person — don't paste templates verbatim. If no, end on the roadmap summary.

If the person scores 0-2 and wants a guided ramp, point them at /quickstart — it onboards them onto the marketing-quickstart framework and gets them to Level 2-3 fast.

---

Self-roast (run before delivering)

• Is the level honest, or inflated? Quantity ≠ maturity — did I check integration + complexity, not just file counts?
• Did I anchor Level 10 to benchmark.md, so the ceiling is real and not flattering?
• Beginner mode: did I explain why each gap matters and define jargon, or did I just list findings?
• Did I show only the NEXT level's roadmap, not all 10?
• Did the dashboard actually render + open, with the per-pillar bars and a shareable card?

---