Getting Started

Core concepts

The building blocks of every ClaudeKit kit — slash commands, skills, agents, shared context files, evidence-first output, and token costs — defined precisely.

Every kit is built from the same handful of primitives. Learn these once and you understand them all.

Kits

A kit is a vertical bundle for one job-to-be-done — engineering, marketing, video, SEO, or e-commerce. A kit ships 16–25 slash commands, 3–5 skills, and 2–4 agents, plus a measured manifest.json, installed with ck install (or as a native Claude Code plugin). Kits do not overlap in scope; you install the one (or ones) that match your work.

KitNamespaceCommandsSkillsAgentsTokens
EngineerKit/eng-*254420,413
MarketingKit/mkt-*203216,714
VideoKit/video-*175312,602
SEOKit/seo-*194216,004
EcomKit/ecom-*203216,464

Slash commands

A slash command is the unit you actually run — a namespaced entry point like /seo-audit or /mkt-repurpose that completes a whole job end to end. Three rules shape every command:

  • Plain language. Descriptions use the words real users use; a non-expert understands every command in one read.
  • Concrete reference material embedded. Commands carry real numbers, templates, checklists, and copy-paste shell commands — never abstract instructions like "optimize the listing."
  • Do everything, never ask. A command finishes the job. If input is missing, it uses the best default and says what was assumed.

Skills

A skill is a dense, focused reference file (SKILL.md) — benchmarks, format specs, voice rules, the exact numbers a command should cite. Skills are what commands reach for mid-run: EcomKit's ecom-benchmarks holds every conversion and email number the kit cites; MarketingKit's mkt-voice-engine holds the anti-AI-tell rules. Each skill's token cost is measured at pack time and shipped in the manifest, so the CLI and your install always agree on what a skill costs. See Token costs.

Agents

An agent is a read-only specialist sub-agent — a reviewer, auditor, or researcher — that a command dispatches for fresh-context work. EngineerKit's code-reviewer reads your diff with eyes that didn't write it; its security-auditor sweeps for injection and secrets. Agents report; they never edit. Keeping them read-only means a command can fan out for evidence without side effects.

Shared context files

Every kit has one shared-context skill that defines the kit's memory files — seo/SITE.md for SEOKit, ecom/STORE.md for EcomKit, VOICE.md for MarketingKit, docs/MAP.md and CLAUDE.md guardrails for EngineerKit. Commands read these files when present and offer to create them when missing — but never block on them. This is how a kit gets smarter across sessions instead of starting cold every time.

Evidence over "done"

Every command ends by showing the result is real: test output, a rendered frame, a before/after diff, a computed table, a verification report. Never a bare "done." This is the quality contract — /eng-verify actually runs your tests and shows the output; /video-make renders the MP4 and checks it; /seo-check prints the exact phrases it flagged. If a command can't prove its work, it tells you that instead of pretending.

Token costs

Your context window is finite, and everything loaded into it costs money on every turn. ClaudeKit treats context as a budget you can measure:

  • Every file in a kit has a measured token cost (ceil(characters / 4)), baked into manifest.json at pack time.
  • The ledger prints on every install, and ck tokens [kit] re-prints the full per-file table sorted by cost — with --live it recounts from the files on disk and shows drift.
  • ck doctor warns when always-loaded context climbs above 25,000 tokens.

This is the ClaudeKit wedge: no other kit ecosystem publishes what its files cost. See the CLI reference for ck tokens.

Lean by design

Kits stay small on purpose. Skill files load their one-line description by default and disclose the full reference only when a task needs it; commands and agents load on demand when invoked. The whole of SEOKit — 19 commands, 4 skills, 2 agents — measures 16,004 tokens end to end, and only a fraction of that sits in context on a given turn. Compared to sprawling free setups that keep tens of thousands of tokens always-loaded, a kit gives you more capability for a fraction of the per-turn spend.