What is actually in the kit
kitstarter is not a bigger pile of skills. It is a small system that changes how your coding agent behaves. Here is every part, what it does to the agent, and the exact misbehavior it kills.
The kit is six working parts: an ask-first engine that makes the agent clarify before it builds, a tutor that teaches one Claude Code power move at the moment you need it, an anti-slop engine that catches AI tells before they ship, a live HUD that shows the agent behaving, sixteen deep skills for the whole build, and a /tour wizard that sets it all up by asking, not lecturing.
The ask-first engine
AGENTS.md plus a hook that will not let the agent guess.
Before any feature-sized request, the agent asks clarifying questions one at a time, in multiple choice, until it could restate your goal with no gaps. Then it reflects the goal back and waits for a yes. In strict mode a hook physically denies file edits until you approve, so it cannot build the wrong thing while you are still deciding.
- The rule lives in AGENTS.md, the cross-tool instruction file, so Codex, Cursor and Antigravity follow it too, not just Claude Code.
- Questions come as 2 to 4 concrete options with the recommended one marked, so you pick instead of writing an answer from a blank page.
- The lock lifts the moment you approve, auto-expires after 30 minutes so a stale flag can never wedge a project, and fails open on any error.
- Say "just do it" and the whole round is skipped. It is your call, always.
This is the fix for an agent that guesses instead of asking, and it takes the teeth out of scope creep and over-building at the same time, because the goal is pinned before a single file is written.
The moment-triggered tutor
The engineered "I did not know it could do that."
The moment you hit a situation Claude Code has a power move for, a feature-sized request, a stubborn bug, a visual bug you are describing in words, a third frustrated retry, the tutor teaches that one move in a few warm lines, then gets out of the way. It never lectures and it never repeats itself.
- Ten lessons, each fires exactly once, ever, tracked in your profile: Plan Mode, extended thinking, checkpoints, screenshots for visual bugs, subagents, connecting MCP tools, context hygiene, resuming work across sessions, automating a repeated ask with hooks, and saving a repeated ritual as your own skill.
- At most one lesson per prompt, and never on a prompt the ask-first engine is already handling, so the two voices never talk over each other.
- The lesson text is plain markdown you can read and edit. No black box.
- One command silences it for good: kitstarter tutor off.
This is the one part that is not about restraining the agent. It is about you: most people use a sliver of Claude Code, and the tutor closes that gap at the exact moment it would have helped. It pairs naturally with the seven failure modes the rest of the kit is built around.
The anti-slop engine
The tells a reviewer spots in two seconds, caught before you ship them.
At wrap-up the agent scans what it actually touched for the giveaways that make code and UI read as AI-made: em-dash prose, glassmorphism by default, tiny fonts, boilerplate no one asked for, fifty lines where one would do. It names each tell instead of quietly shipping it.
- Three levels, your call: notify names the open tells at wrap-up and lets you ship anyway (default); strict will not let the agent finish while a strong tell is open; off turns the scan off.
- A subagent never inherits the session's rules, so a SubagentStart hook re-injects "stay lean, no slop, report straight" into every subagent the agent spawns, closing a hole other rule-kits leave wide open.
- The /lean and /polish skills fix what the scan finds by the root cause, not by decorating over it.
This is the fix for shipping code that looks AI-made, and it is where the kit earns its own standard: the product that fights slop cannot ship any.
The live HUD
kitstarter watch: see the agent behaving, do not just trust it.
Run kitstarter watch in a second terminal pane and you get a live dashboard beside your agent: your project roadmap with progress, the clarity and tutor status lights, and an activity feed that names every move as it happens, goal confirmed, edit blocked, lesson taught. You watch the behavior instead of hoping for it.
- The bot-2 mascot is rendered right there in the terminal as truecolor pixel art, and it types when the agent is busy, blinks when it is idle.
- The feed reads the same events the hooks write, so a blocked edit or a confirmed goal shows up the instant it happens.
- Plain ANSI, no heavy TUI library. Works in Windows Terminal, macOS Terminal and iTerm2, and Linux.
It is the answer to an agent that says done but never checked: when you can see whether it confirmed the goal and whether an edit was blocked, "done" stops being a thing you take on faith.
- scaffold the app
- recipe list + detail
- add search
- save a recipe (auth)← here
- deploy
Sixteen skills, each one deep
A set you will actually use, mapped to the arc of a real build.
Not 108 commands you will never remember. Sixteen skills that cover a build end to end: plan it, build it a piece at a time, prove it, ship it, rescue it when it breaks, and understand what you made. Each one is a full playbook, not a one-liner.
- They are behavior, not capability bloat: /start and /review encode the same ask-first, prove-it discipline the hooks enforce, so the whole kit pulls in one direction.
- You do not have to memorize them. Type the slash command, or just describe what you need, and the matching skill recognizes the moment and triggers itself.
- Cross-tool where it can be: the skills run in Claude Code; the ask-first rule underneath them travels to any agent that reads AGENTS.md.
Between them they cover all seven ways an agent goes wrong, from over-building to a leaked key on the day you go live.
/startinterrogate the idea until the goal is clear, then write the plan/stepbuild one piece of the plan, then save your place
/checkprove a change really works before calling it done/reviewfresh eyes on the diff against the original goal/shipget it live, set env vars, hand back a URL
/fixfind the real cause before changing anything/recoverget back to when it worked, calmly, without git fear/secretscatch a leaked key or exposed .env before it goes public
/leancut the over-building and one-use abstractions/design-intakedecide the look on purpose, write a DESIGN.md/design-checkan honest read on whether a screen looks AI-made/polishfix what makes it look generated, by the root
/explaina plain-English readout of what the AI actually built/claude-tipsthe Claude Code features you are not using yet/connect-mcpgive the agent real tools: web, live docs, a browser/tourthe first-run wizard that sets the whole kit up
The /tour setup wizard
The product demonstrating itself on the first run.
The first thing you run. Instead of a wall of documentation, /tour is kitstarter being kitstarter: it equips your agent with the essential opt-in tools, then onboards you by asking one question at a time, in multiple choice. Two minutes and it is tuned to your project, not a generic default.
- npx kitstarter-cli installs everything in one command; add --global to arm every project you open.
- It asks, it does not lecture, which is the whole point: your first taste of the tool is the tool's core move.
- Change your mind later with plain commands: kitstarter clarity strict|relaxed|off, tutor on|off, slop notify|strict|off, and kitstarter remove for a clean uninstall that keeps your files.
It is why a total beginner can get value on day one. If you are still deciding whether that is you, read who the kit is for.
The things you are already thinking
Fair questions. Here are the straight answers before you ask them.
Can't I just write AGENTS.md myself?
Yes, and you should. The free ask-first rule is the whole idea, and we give it away on the fix pages. But a markdown file is a polite request the agent forgets by the third prompt. The kit is the part that makes it stick: a hook that denies the edit until you approve, the slop scanner, the tutor, the HUD. The rule is the idea; the kit is the enforcement.
Isn't this just another pile of skills?
It is the opposite bet. The category is racing toward "108 skills, 16 agents" and toward free. One near-clone ships 18 skills plus 16 agents at $89. kitstarter ships 16 skills, each deep, plus the behavior engine a pile of commands does not give you. You are buying how the agent acts, not how many commands it has.
Won't this lock me to Claude Code?
No. The ask-first rule lives in AGENTS.md, which Codex, Cursor and Antigravity all read, so the behavior travels with you. The full engine, the clarity lock, the tutor, the HUD, the slop gate, runs in Claude Code today, where most of the pain actually is.
Why pay when Anthropic's marketplace is free?
Because those are capability plugins: they add things the agent can do. kitstarter changes how it behaves, which no plugin does. Different axis entirely. Use both.
$29. Is it actually worth it?
It is one payment with lifetime updates, less than an hour of the rework a guessing agent costs you. Try the free ask-first rule first. If your next session does not feel different, do not buy the kit.
Common questions
How do I install it, and what does it touch?
One command: npx kitstarter-cli. It wires the hooks into your Claude Code settings and drops the 16 skills in place, then runs /tour to set up. Add --global to arm every project you open. kitstarter remove unwires it cleanly and keeps your own files.
Will the hooks slow me down or block me when I do not want them to?
The clarity lock only engages in strict mode, only on vague build requests, and it lifts the moment you approve. "Just do it" skips it entirely, it auto-expires after 30 minutes, and it fails open on any error, so it can never wedge your session. Planning files stay writable throughout.
Can I turn parts off?
Yes, all of it is tunable with plain commands: kitstarter clarity strict|relaxed|off, kitstarter tutor on|off, and kitstarter slop notify|strict|off. Set it once in /tour, change it any time.
Which agents and operating systems does it work with?
The full engine runs in Claude Code on macOS, Windows and Linux. The ask-first rule itself works in any agent that reads AGENTS.md, including Codex, Cursor and Antigravity.
Is it a subscription?
No. It is a one-time payment with lifetime updates. $29 is the kit; $59 adds a single 1:1 coaching session. Same software, optional human help, not bigger software.
Make your agent ask before it builds
One kit. Ask-first behavior, lean output, no AI slop. Works with Claude Code, Codex, and Antigravity.
Get the kit · $20 $29One-time. Less than one hour of cleaning up AI slop.