██████╗ ██╗ ██╗████████╗███████╗
██╔══██╗██║ ██║╚══██╔══╝╚══███╔╝
██████╔╝██║ ██║ ██║ ███╔╝
██╔══██╗██║ ██║ ██║ ███╔╝
██████╔╝███████╗██║ ██║ ███████╗
╚═════╝ ╚══════╝╚═╝ ╚═╝ ╚══════╝
⚡ A holistic-machine Claude Code plugin for Vue/Nuxt + Firebase ⚡
37 skills · 11 agents · 38 hook scripts across 16 events · 13 shared protocol files
Orchestrator main-thread router · 7 anti-shortcut hooks · 8-invariant quality ratchet · optional Cross-Model Critic
What is Blitz?
Blitz is a holistic machine: a set of parts wired so natural-language intent goes in one end and shippable, gated code comes out the other — with no step able to silently skip the one after it. It turns Claude Code into an opinionated, partly-autonomous development environment for Vue/Nuxt + Firebase.
The mechanism in one breath: freeform input lands on a read-only orchestrator that routes to a /blitz:* skill; skills spawn worker agents in isolated worktrees; every artifact they produce is checked at a gate that runs off a single shared rule registry; an adversarial critic must sign off before anything reaches PASS; and the whole loop persists its state to disk so it can resume itself unattended.
flowchart LR
U([Freeform intent]) --> O[Orchestrator<br/>read-only router]
CMD([/blitz:skill/]) --> SK
O -- routes to --> SK[Skill]
SK -- spawns --> AG[Worker agents<br/>isolated worktrees]
AG --> G[Gate<br/>shared check-registry]
G --> C{Critic<br/>LGTM?}
C -- REJECT --> SK
C -- LGTM --> P([PASS / ship])
G -. writes state .-> ST[(.cc-sessions/)]
ST -. resume / re-tick .-> O
Four properties make it a machine rather than a pile of prompts:
- State is external and append-only. Scope claims, sprint progress, quality metrics, and cross-session handoff all live in
.cc-sessions/ — so any tick can reconstruct what the last one was doing.
- Gates run on data, not vibes. Detection logic lives in one
check-registry.json; skills and critics select from it. A check either fires deterministically (grep/tsc/git) or is an LLM judgment — and the registry records which, so only facts can block.
- Verification is adversarial and external. The critic's job is to find one reason to REJECT; the research-critic's job is to catch a hallucinated citation. Both are read-only and can be routed to a different model family to dodge home-model blind spots.
- It closes the loop.
/blitz:next --loop reads state, dispatches exactly one phase, commits, and exits — so /loop or a scheduler can re-tick it toward "shippable" without a human in the chair.
Reading Tiers
Quick Start
npx blitz-cc@latest
The installer detects your stack, registers the plugin, configures permissions, and wires the hooks. From then on you mostly type intent and let the orchestrator route.
More install options
Non-interactive: npx blitz-cc@latest --yes
Bash fallback (no Node): curl -fsSL https://raw.githubusercontent.com/lasswellt/blitz-cc/main/installer/install.sh | bash
Marketplace: /plugin marketplace add lasswellt/blitz-cc then /plugin install blitz@blitz
Local dev: claude --plugin-dir ./blitz then /reload-plugins
First five minutes
/blitz:next # reads state, recommends the next command
/blitz:ask "I want to add X" # routes a vague request to the right skill(s)
/blitz:research <topic> # parallel-agent investigation → docs/_research/
/blitz:sprint # full cycle: plan → implement → review
/blitz:review # the precision quality gate, on demand
/blitz:ship # gates → release → publish
Or just type freeform — the orchestrator routes for you.
Prerequisites
