Sailop vs v0, Bolt, Lovable, Cursor and Claude Code: Where Each Default Comes From

Every AI coding tool falls back to a default when the prompt leaves room. The useful comparison is not which tool scores best on a blank prompt, but where each default comes from and what each tool gives you to override it. Those two things are documented, so this page compares them, then says plainly what Sailop does and does not do today.

An earlier version of this page reported a side-by-side test: six tools, one prompt, three runs each, per-dimension scores. We hold no record of that test, and two of the tools in it (the Sailop CLI and its composer) are not published, so nobody could have reproduced it. It has been removed. Everything below comes from each vendor's documentation, read on 17 October 2026, or from Sailop's own saved scans, which are dated.

Where each tool's default comes from

The first three generate a project from nothing, so their starting point is the tool's own. The last two work inside your repository, so their starting point is your code; on an empty folder, they fall back on the same habits as everyone else, because the models learned from the same web.

What each tool gives you to override the default

v0

• Design Systems 2.0: import your components, tokens and a starter app once; v0 saves them as a skill that new chats can start from. On paid plans, a team owner can set a default design system for new chats.
• Instructions: custom instructions applied to v0's responses.
• Design Mode: select an element in the live preview, change its styles, apply the edit back to the code.

The catch, in v0's own registry guide: it "may struggle" with customized versions of shadcn/ui components, because it was trained on the defaults. Make primitive-level changes after generation, or give it a design system so it is not guessing.

Bolt.new

• Project knowledge, which Bolt calls "persistent": the model can always access it, even after the chat context is cleared.
• An agents.md file: "Bolt finds and uses it automatically."
• Design systems on Teams plans, and Figma import as design reference.
• Visual edits in the preview, saved to code.

Lovable

• Design guidance: when a prompt leaves the look open, Lovable asks about palette, typography and layout before it builds. Answer them, or name the look in the prompt and it "builds directly."
• Knowledge, per project and per workspace, for rules that apply to every chat.
• Design systems and Figma import (a plugin, the Figma MCP, or a .fig upload).

Cursor

• Project rules in .cursor/rules, "version-controlled and scoped to your codebase," which can apply to every chat, to matching files, when the agent judges them relevant, or only when mentioned.
• AGENTS.md at the root or in subfolders.

Claude Code

• CLAUDE.md files at project, user or organization level, loaded at the start of every session, and .claude/rules/ for rules scoped to file types.
• Its documentation adds a caveat that applies to every tool on this page: Claude "treats them as context, not enforced configuration." Rules steer; they do not guarantee.

What this comparison actually shows

1. Constraint choice matters more than tool choice. Every tool above has a place to put your design decisions. None of them invents those decisions for you. A specific typeface, a palette you chose, a layout that is not the average, and real copy, written down where the tool reads them, do more than switching tools.

2. Builders and editors differ in where the constraint lives. In v0, Bolt and Lovable it lives in the platform: a design system, a knowledge entry, answers to design questions. In Cursor and Claude Code it lives in your repository, next to the code, under version control. If you work across several tools, the repository is the one place they can all read.

3. Rules are context, not enforcement. The Claude Code docs say it outright and the others imply it. Check the output after each change, especially after an unrelated request.

What Sailop does today

One thing: a free web scan. Give it the address of a live page and it checks the page's markup and stylesheets for eight signatures of AI-default design:

The score is the sum of the weights of the signatures found, capped at 100, and the result comes with a fix kit: a palette and type pairing seeded on the site's own name, find-and-replace swaps built from the page's tokens, and a brief to paste into whichever coding tool you use.

What it measures is the presence of defaults, not taste. Three public pages scanned on 13 October 2026 show the range:

tailwindcss.com carries seven of the eight signatures, which is no criticism of Tailwind's own site: it showcases its framework's own utilities. flowbite.com carries four. stripe.com carries one, frosted glass. A high score says "these are the defaults everyone ships," and the fix kit says what to change.

What Sailop does not do today: the sailop command-line tool and MCP server described on parts of this site are not published on npm yet, so they cannot be installed, and nothing on this page depends on them. The install page tracks their status.

How to get less generic output from any tool

• Decide before you prompt. Two typefaces, two or three colors, one reference site that is not a SaaS landing page, and the layout of the first screen.
• Write it where the tool reads it. A v0 design system, Bolt's project knowledge or agents.md, Lovable's knowledge or design answers, Cursor's .cursor/rules, Claude Code's CLAUDE.md.
• Ask for structure, not adjectives. "Asymmetric hero, text left, product screenshot cropped off the right edge" beats "modern and clean."
• Scan the result. Run the free scan on the preview URL or the deployed page and work through what it finds.
• Fix by hand what the model keeps undoing. The defaults return under pressure; the last pass is yours.

For the underlying mechanism, see why every AI-generated website looks the same. For the patterns themselves, the card-grid problem and the 10 dead giveaways. For the tools one by one: v0, Bolt.new, Lovable.

Disclosure

Sailop competes, at the edges, with the tools above, and the scanner described here is ours. The scan is free and its eight checks are listed in full above, so you can judge what it measures. The three sample results are saved exactly as the scanner returned them on 13 October 2026.

Revision note

Rewritten on 17 October 2026. The earlier version, "Honest Side-by-Side Test (April 2026)," reported per-tool scores, per-dimension breakdowns, line counts and before-and-after results for Cursor and Claude Code from a test we hold no record of, and promoted a CLI and paid tiers that cannot be bought or installed. All of that has been removed. The page now compares what each tool documents about its defaults and its overrides, and describes the free scan as it works today.