It guesses instead of asking
You ask for a login and it builds an entire auth framework. Vanilla agents assume what you meant and run with it, so you get the wrong thing fast.
The MUST-HAVE starter kit for Claude Code and CodexThe AI coding starter kit
The AI coding starter kit
that asks before it builds.
Vanilla agents guess and over-build. kitstarter makes Claude Code and Codex ask the right questions, build lean, and ship a real product that doesn't look AI-made. You learn the whole workflow, from planning to deployment.
Make Claude Code and Codex ask the right questions, build lean, and ship work that doesn't look AI-made.
Get the kit · $25$50One-time. Installs in a minute. Works in your terminal.$25 once. Installs in a minute. Full refund if it doesn't help.
voice
›
kitstarter is reading your request…
New here? Don't get left behind.
Never built anything? Start now, before AI leaves you behind.
The gap between people who can build with AI and everyone else grows every week. If you're a total beginner, sign up and I'll personally walk you through the whole thing, from install to your first live project. Free, and the only wrong move is waiting.
Get the kit · $25$50What's in the kit
$ npx kitstarter-cli
✓ 16 skills installed
/start/step/check/fix/review/ship/secrets/recover/polish+7 more
- Asks before it builds
- No 40-file surprises
- Catches the AI-made look
- Proves it runs before it says done
Tired of wrestling with your AI agent?
Claude Code and Codex are powerful. Getting consistent, production-quality work out of them is the frustrating part. Run them vanilla, and this is what you fight.
It forgets your project
Context runs out and the agent loses the plot. Every new session it reinvents your conventions and contradicts last week's code.
It over-builds everything
A one-line fix turns into a new dependency and forty files. You inherit code you never asked for and now have to maintain it.
It ships AI slop
Gradient buttons, emoji headings, generic copy. The output screams AI made this, and your users can tell.
You redo the same setup
Copy-pasting prompts between projects and rebuilding the same patterns, with no system that actually sticks.
It says done without checking
It calls the work finished without running the tests. You find out it is broken when your users do.
Same AI. Two completely different outcomes.
The kit doesn't swap your model. It changes how it behaves, from guessing and over-building to asking and shipping clean.
Guesses, then over-builds
- Assumes what you meant and builds the wrong thing, fast
- 40-file diffs; a library for what a line would do
- Violet gradients, generic font, looks obviously AI-made
- Says “done” without ever running it
✦ With kitstarter
Asks first, stays lean, ships clean
- ✓ Asks the right questions before it touches code
- ✓ Smallest change that actually works, no bloat
- ✓ Catches the AI-slop tells, looks intentional
- ✓ Proves it runs before it calls it done
One workflow, planning to deployment.
Your agent follows a method that ships an actual product and teaches you how it works along the way.
01
Understand
It asks the questions that matter before it writes a single line.
02
Plan
A short, real plan you approve, not a 40-file guess.
03
Build lean
One step at a time: the smallest change that actually works.
04
Prove it
Runs the tests and shows you green before it moves on.
05
Ship & deploy
Catches leaked keys, gets it live, and recovers when it breaks.
What's inside. Few skills, each deep.
Not 108 commands you'll never learn. A few skills, each deep, and one command installs all of it.
always on
The core
Ask first, be direct. The always-on rule, plus the setup wizard.
- /start
- /tour
- the AGENTS.md rule
- session greeting
Build cleanly
One step at a time, prove it works, cut the over-built code.
- /step
- /check
- /fix
- /review
- /lean
Ship & survive
Get it live, catch leaked keys, and recover when it breaks.
- /ship
- /secrets
- /recover
- /explain
37-tell detector
Don't look AI-made
A real detector for the tells that scream “AI generated this”, and the skills to fix them.
- /design-intake
- /design-check
- /polish
Reach further
Web, search, and live docs (crawl4ai + exa), plus the Claude Code and Codex you're not using.
- /connect-mcp
- /claude-tips
opt-in
Voices
Switch to a teaching or buddy tone when you want it. Default stays direct.
- teaching
- buddy
16 skills.
One that matters most: it asks.
One that matters most: it asks.
Drop it into Claude Code or Codex and the whole engine runs natively in both: the skills, the slop detector, and the hooks. The ask-first rule works anywhere that reads AGENTS.md, including Antigravity.
Get the kit · $25$50You're already building with AI. Build it right.
Three kinds of people buy the kit. Find yourself.
You just installed Claude Code
Your first builds work until they suddenly don't. The kit makes the agent ask what you don't know to ask, so you skip the collapse everyone else learns the hard way.
Start right, today
You still write every line by hand
Directing agents is how teams ship now. Learn to run one properly before it's the thing your next role assumes you know.
Level up now
You run a business, not a codebase
The tool your business needs could be live this week. The kit keeps the agent on rails so what you ship actually works.
Build it this weekOne-time. Yours forever.
Launch offer$25 one-time, normally $50.
Best-seller
The kit
$25$50once
Less than one hour of cleaning up AI slop.
Everything you need to make your agent ask first and build clean.
- ✓ All 16 skills plus the 37-tell slop detector
- ✓ Auto hooks: anti-slop, safety, the greeting
- ✓ Teaching and buddy voices, direct by default
- ✓ Lifetime updates as Claude Code moves
Full refund if it doesn't calm your builds.
One plan, one price worldwide, in US dollars. One-time, no subscription.
Questions, answered.
What is kitstarter?
kitstarter is a one-time behavior kit for Claude Code and Codex that makes your AI coding agent ask clarifying questions before it builds, stay lean, and stop looking AI-made. You get durable project guidance, reusable skills, and native lifecycle hooks in both tools. It installs with one command: npx kitstarter-cli.
How do I make Claude Code ask questions before it writes code?
Install kitstarter with npx kitstarter-cli. Its always-on rule makes the agent ask the questions that matter before it touches code, instead of guessing what you meant and over-building.
Does kitstarter work with Codex?
Yes, natively, and here is the honest shape of it. Codex gets 11 handlers across 7 lifecycle events, wired through .codex/hooks.json: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, PreCompact, SubagentStart, and Stop. The guard, the clarity check, the slop detector, the live HUD, and the tutor (with its own Codex lessons) all run there, not just an AGENTS.md file. The wiring is Codex-native, it matches Codex's own apply_patch tool, and it merges into your existing hooks.json rather than overwriting it. Two real differences from Claude Code, because we would rather you heard them from us: Codex publishes no SessionEnd event, so the end-of-session recap is Claude Code only, and Codex's PreToolUse is warning-only, so where Claude Code can hard-deny a risky edit, Codex gets a strong instruction instead.
Does kitstarter work with Cursor or Antigravity?
The ask-first rule works in any agent that reads AGENTS.md, which includes Antigravity and Cursor. The full stateful hook engine (the guard, the slop detector, and the tutor) runs natively in Claude Code and Codex today.
How much does kitstarter cost?
It is a one-time payment, no subscription: $25 (normally $50). One plan, the kit, and it includes everything: all the skills, the 37-tell slop detector, hooks, voices, and lifetime updates.
How do I install kitstarter?
One command in any project: npx kitstarter-cli, which installs everything in one tier. Add --global (or -g) to install it for every project at once.
What is AI slop, and how does kitstarter stop it?
AI slop is output that screams that AI made it: gradient buttons, emoji headings, generic copy, over-built code. kitstarter ships a 37-tell detector and design skills that catch those patterns so your work looks intentional.
How is kitstarter different from other AI agent kits?
Those kits sell more agents and skills. kitstarter changes how your agent behaves: it asks before it builds, writes the smallest thing that works, proves it runs, and stops the output looking AI-made. It is about behavior and craft, not volume.