The honest way to compare vibe coding tools is to read their agent’s rulebook

The actual lever, drawn out

Here is the path a single user prompt walks before the agent starts writing. The model is the bottom of the stack. The rules are the top.

Three of those five steps are platform decisions. You only control the first one, the user prompt, and only sometimes the third, the project CLAUDE.md. The platform owns the system prompt. The platform also owns whatever skill files the agent has access to. So when you compare tools, you are comparing the platform’s taste, not just its model choice.

Read those again slowly. The rule at line 84 collapses the color choices the agent will reach for from a palette of dozens to a palette of three. The rule at line 126 to 127 rules out the most recognizable AI font signature on the open web. The list at line 152 names, by name, the single most common AI app maker output pattern: hero section centered, gradient button, purple-to-indigo, soft shadow, rounded card grid below. That output is a habit, not a choice, and habits are killed by writing them down as anti-patterns and feeding them to the agent on every turn.

# docker/e2b/e2b.Dockerfile, around line 89

# ── Claude Code config at /root/.claude ──
COPY files/root/.claude/ /root/.claude/

# Result: every sandbox the agent boots into has the
# rulebook on disk before it answers your first prompt.

The five point checklist you can run on any tool

You do not need insider access to apply this lens. Here is a test you can run on any vibe coding tool today, including mk0r. The point is not that mk0r passes all five. The point is that almost every other roundup ignores all five.

1. Is the system prompt visible. Can you read what the agent has been told before you opened the chat. If the answer is no, you are renting the platform’s taste sight unseen.
2. Is there a project rules file the agent reads on every turn. CLAUDE.md, AGENTS.md, .cursorrules, agent.md. Different names, same idea. If there is none, you cannot enforce taste across multiple turns.
3. Does the rulebook name anti-patterns by name. Generic guidance like “produce beautiful code” is decorative. Specific bans (no Inter, no purple-on-white, three colors max) are load bearing.
4. Are the rules in source control where you can read them. A platform that publishes its agent’s rulebook is making a claim it has to live with. A platform that hides it is asking you to trust the marketing copy instead.
5. Does the output actually obey the rules. This is the empirical leg. Build the same prompt on three tools. The fonts, colors, and section shapes you get back are the de facto system prompt, whatever the marketing says.

The same prompt, on a tool with rules and a tool without

To make the difference legible, here is the kind of output you can expect when the agent has no design rulebook in the loop versus when it does. The user prompt is identical in both cases: “a one page habit tracker, single daily habit, light theme, no auth, no database.”

The grid: what most lists rank vs. what actually changes the output

Every existing best-of roundup grades vibe coding tools on features. The features matter. They are not what changes the shape of your output. Here is the side by side.

The counterargument I want to take seriously

The reasonable pushback on all of this is: the model is the model, and a strong enough model will produce strong output regardless of what you put in the system prompt. There is something to that. A frontier model is much harder to steer into AI slop than a weaker one. But the empirical evidence on the open web cuts the other way. The most common complaint about AI generated apps in 2026 is not that the code is wrong; it is that everything looks the same. That uniformity is the model defaulting, not the model failing. A rulebook in front of the model breaks the default.

The other reasonable pushback is: rules constrain creativity. They do. The honest tradeoff is that without them, the output is creative on a single dimension (the words you typed) and absolutely conformist on every other dimension. With them, the output is constrained on the dimensions where the model would otherwise default to conformity, and free on the dimensions you actually care about. That is the trade I think most makers want, and it is the trade mk0r’s rulebook makes on your behalf.

Where mk0r lands on the checklist

To be fair, here is mk0r against the same five point test I said you should run on any tool, including this one. I am marking the cells honestly. Some of these are stronger than others.

The honest take on the broader landscape

mk0r is not the right pick for everyone. If your work flows through an IDE and you want an AI pair sitting in your editor, Cursor or Windsurf will fit better than any of the web app builders. If your project is a back-office tool that needs role-based access, a relational schema, and form-driven CRUD pages, a tool that ships those primitives will save you weeks. If you are a non-developer who wants a full-stack app with the lightest possible learning curve, Lovable, Bolt, or Replit Agent are all reasonable picks with their own tradeoffs.

The argument here is narrower. It is that whichever tool you pick, the most predictive single thing about what your generated app will look like is what the agent reads on every turn. Most lists will not tell you that, because opening the rulebook is more work than counting features. The work is worth doing. It is also the only way to know why two tools using the same model produce wildly different output, and why a third tool ranks last on every feature chart but produces the only output that does not look like everyone else’s.

One small experiment to try this week

Pick three vibe coding tools you are considering. Open all three in different tabs. Type the same prompt into each: a single page habit tracker, one daily habit, light theme, no auth, no database. Do not say anything about fonts, colors, or layout. Screenshot the first preview each tool produces. Lay the three screenshots side by side.

Two of them will look like siblings. The fonts will read as Inter. The CTA will be a violet or indigo gradient. The layout will be a centered hero on top of a uniform card grid. That is the AI default coming through unfiltered. The third, if you are lucky, will be visibly different. If none of the three is different, none of the three has a rulebook in the loop, and the right answer is to keep looking.

Then run the same prompt on mk0r with no signup. The defaults you get back are not perfect. They are obviously not Inter, not a violet gradient, and not a centered hero with a card grid below. That is the rulebook earning its keep, line by line.