hypt — Shipping Workflow for AI Coding Agents

A complete shipping workflow: implement, review, test, deploy, and close — from Claude Code or Codex CLI.

Get hypt running (no terminal needed)

If you're new to coding, start here. The whole flow takes about 5 minutes and you don't need to know any commands.

1. Install Claude

Download Claude (Mac and Windows). It's free with an Anthropic account — sign-up takes about 30 seconds.

On Linux, or already have a terminal you like? Skip to Manual install (for developers).

2. Open the Code workspace

In Claude, click the Code option in the sidebar. (Or, in a browser, go to claude.ai/code.) This is the part of Claude that can install tools and edit files on your computer.

3. Tell Claude to install hypt

Paste this into the chat exactly as written and press Enter:

Install this plugin: https://github.com/brianevanmiller/hypt-builder

Claude will:

• Check what's already on your computer (Git, Node, Bun, GitHub CLI)
• Ask before installing anything that's missing — usually one or two clicks of "yes"
• Set up hypt itself

You may see a panel below the chat where Claude is running commands. You don't need to type anything there yourself — just answer Claude's questions in the chat above. If your computer asks for your password, that's normal — it's how your operating system installs developer tools safely.

This usually takes 1–2 minutes.

4. Start your project

When Claude tells you "hypt is installed," paste:

/start

That's it. From here, /start walks you through:

• Describing your app idea, in plain English
• Signing up for the services your app needs (GitHub, Vercel, Supabase, plus optionally Stripe / Resend / a custom domain) — each one has a quick web signup that /start opens for you
• Scaffolding your project and writing a build plan

You don't need to set up any of those services in advance. The whole thing is designed for someone who has never opened a terminal.

---

Manual install (for developers)

Already have Claude Code or Codex CLI installed? Skip the desktop app:

macOS / Linux:

bash <(curl -fsSL https://raw.githubusercontent.com/brianevanmiller/hypt-builder/main/install.sh)

Windows (PowerShell):

irm https://raw.githubusercontent.com/brianevanmiller/hypt-builder/main/install.ps1 | iex

After installation, restart your agent and run /start.

The installer expects git, node (>=18), bun, and gh to already be on your $PATH. If anything's missing, install it via Homebrew (brew install …), winget, or your distro's package manager — or just use the Get hypt running flow above and let Claude handle it.

Starter CLAUDE.md

During installation, hypt offers to install a starter ~/.claude/CLAUDE.md file. This gives Claude built-in engineering discipline — planning before building, verifying work, learning from corrections, and following git best practices. It's optional and fully customizable.

If you already have a CLAUDE.md, hypt offers to enhance it by appending the engineering discipline section — your existing content is preserved. The starter is always available at docs/starter-claude-md.md for reference.

Codex CLI Support

Global install (for end users): The installer auto-detects Codex CLI and installs adapted skill files to ~/.hypt/skills/ with a global instruction block in ~/.codex/instructions.md. Skills are available via natural language (e.g., "save my changes", "review my code").

Repo-native (for hypt contributors): This repo also ships repo-native Codex skills under .codex/skills/ with an index in AGENTS.md. To regenerate after changing Claude sources:

node scripts/sync-codex-support.mjs

Cheatsheet

New to hypt? See CHEATSHEET.md for a printable one-page reference with the commands you'll use most. After running /start, you'll be offered the cheatsheet automatically.

Commands

All commands use the hypt: prefix (e.g., type /start or /hypt and Claude will route to the right one):

Command	Description
/start	Onboarding — describe your idea, set up accounts, create a build plan
/prototype	End-to-end: review plan, implement, review x2, test, and deliver
/save	Commit, push, and create/update PR automatically
/review	Thorough PR review with 4 parallel subagents — auto-fixes urgent issues
/touchup	Quick pre-merge polish — fix PR comments, build issues, update docs
/unit-tests	Smart unit test generation prioritized by business criticality
/fix	Diagnose and fix bugs — triage, research, plan, and deliver a tested fix
/deploy	Verify deployment health — detects platform, auto-bypasses Vercel free-plan blocks
/status	Quick deployment status check — is my site up?
/restore	Restore to a previous working version — rollback, revert, database recovery, auto-investigates
/post-mortem	Analyze what went wrong after a restore — creates incident doc, updates backlog
/docs	Scan and update project documentation — checklists, READMEs, feature docs
/close	Suggest next tasks, update backlog, confirm before merge, verify deployment
/todo	Add or update items in your project's backlog, roadmap, or todo list
/suggestions	Suggest next tasks, group related items, and offer /yolo or /go activation
/plan-critic	Critical plan review — find gaps, ask questions, refine before building
/go	Autonomous pipeline + confirm before merge
/yolo	Fully autonomous pipeline + merge, no stopping
/pipeline	Full development pipeline — research, plan, build, review, test, save PR (no merge)
/autoclose	Autonomous merge — deploy check, version bump, release (no confirmation)
/ci-setup	Set up lightweight CI — runs unit tests on every commit

Workflow

Just type what you want — the router picks the right skill:

 ┌──────────────────┐     ┌────────────────┐     ┌──────────────┐
 │  "fix this bug"  │────►│  hypt router   │────►│   hypt:fix   │
 │  "ship it"       │     │ (phrase match)  │     │   hypt:close │
 │  "save"          │     └────────────────┘     │   hypt:save  │
 └──────────────────┘                            └──────────────┘

The typical development flow:

 start ─► prototype ─► save ─► review ─► touchup ─► tests ─► docs ─► deploy ─► close

Shortcuts compose skills into full pipelines:

 /go   = pipeline ─► confirm ─► autoclose    (autonomous with safety net)
 /yolo = pipeline ─► autoclose               (fully autonomous)

Each command can also be used independently. See docs/hypt-router-design.md for detailed routing diagrams and skill composition.

gstack Integration (Optional)

hypt works great on its own, but for the full experience, install gstack — a free companion tool that adds 35+ specialist skills:

• Visual QA testing (/qa) — test your app in a real browser
• Design review (/design-review) — spot visual issues and AI design slop
• Security audit (/cso) — OWASP Top 10 + STRIDE threat model
• Product thinking (/office-hours) — YC-style forcing questions to sharpen your idea
• Systematic debugging (/investigate) — root-cause analysis for complex bugs
• Design exploration (/design-shotgun) — generate AI mockup variants
• Performance testing (/benchmark) — Core Web Vitals and page load times

When gstack is installed, hypt automatically detects it and:

• Routes QA, design, and security requests to gstack's specialist skills
• Escalates complex code reviews and bug investigations to gstack's deeper tools
• Offers product thinking via /office-hours during /start onboarding

Install gstack:

git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup

Or just tell your AI agent: "Install gstack"

Without gstack, hypt handles everything itself — code review, testing, deployment, and bug fixes. gstack adds depth for visual testing, design, and security that hypt doesn't cover natively.

Matt Pocock's Companion Skills (Optional)

Matt Pocock (creator of AI Hero and Total TypeScript) maintains a skills repo with focused, single-purpose tools that pair well with hypt. /start will offer to install these for you, but you can also install them manually:

• /grill-me — gets you relentlessly interviewed about your plan, one question at a time, until every decision is nailed down. Pairs with /plan-critic (which is more analytical) — /grill-me is the conversational version.

  bunx skills@latest add mattpocock/skills/grill-me -g -y

• git-guardrails — installs a Claude Code hook that blocks dangerous git commands (push --force, reset --hard, clean -f, branch -D) before they can run. Belt-and-suspenders safety for non-coders.

  bunx skills@latest add mattpocock/skills/git-guardrails-claude-code -g -y

These skills install via Vercel Labs' skills CLI, a universal installer that works with Claude Code, Codex, Cursor, and ~50 other agents.

Why not the rest of Matt's skills? Matt also publishes /to-prd and /to-issues (turn conversation into a PRD, break a plan into tasks) — but they hardcode GitHub Issues as the output. hypt prefers docs/ files and docs/todos/backlog.md for tracking, so we don't recommend those by default. A hypt-native equivalent that writes to docs/ (and optionally syncs to Linear / Notion / your tracker of choice) is sketched in docs/linear-integration-plan.md.

Supported Deployment Platforms

Vercel, Netlify, Fly.io, Render, Railway, and GitHub Deployments API (fallback).

Security

Includes a supply chain security scanner that runs in CI on every PR to main. Detects prompt injection, invisible Unicode attacks, shell injection, tool poisoning, and structural anomalies. See docs/security-scan.md for details.

Requirements

• Node.js — required by Claude Code and used by the installer
• GitHub CLI (gh)
• Git
• Bun — runtime, package manager, and task runner (needed for /start and /prototype, not for install)

License

MIT