StyleSeed: turning design rules into a gate agents cannot ignore

A rules engine, not a prompt dump

StyleSeed turns recorded design decisions into project-local rules and implementation workflows, and the subheading under the title makes its stance clear: it supports experts rather than replacing them with a collection of tastes.

The pitch is a single claim about the same product and the same prompt. Only the rules changed. The README shows a comparison image with default AI output on one side and StyleSeed rules locked on the other, and a second animation of one chat UI morphing across brand skins.

The measured result is stated with a scope: on 120 rendered benchmark cells, the enforced gate improved both Codex and Claude Code by 5.3 points. And the more interesting number follows: raw rules alone were inconsistent, with Codex at plus 1.6 and Claude Code at minus 3.7. That negative result is the reason the project ships a render, score, revise loop instead of a checklist, because the README argues that blocking the obvious defaults is not enough, since agents converge on new defaults once the old ones are removed.

So the design claim is not that rules fix the output. It is that rules need an enforcement loop, and the enforcement is a gate.

The project is MIT licensed, the primary language is JavaScript, the homepage is a demo site at styleseed-demo.vercel.app, and the last push to the main branch was on 2026-10-14, with v5.0.0 published on 2026-09-30 and v4.2.0 on 2026-09-11.

The install, and the thirty second path

The install is one command, and the README is explicit about what it does and does not do.

npx -y skills add bitjaru/styleseed -a codex claude-code -y --copy

It installs the core skills for Codex and Claude Code together, with no agent, skill or install-mode menus, and it copies the skills into this project. It does not install globally and it does not change other projects, which is the right default for a set of project-local rules.

The -y flag skips the installer questions, and the README is careful to add that it does not bypass your coding agent's tool permissions.

Then you send one message. For Codex it is $styleseed Help improve this screen. Preserve our existing design choices and show me what needs a decision, and for Claude Code the same text with a slash command instead of a dollar sign. Only styleseed is registered, and it selects the needed internal workflow from your request.

The workflow the README describes is a conversation rather than an automatic pass. You choose the unresolved design decisions together, then you inspect the actual rendered result, which is why the second step is asking the agent to show you what needs a decision rather than just to change things.

The collapsed section covers the variants. Codex only is -a codex, Claude Code only is -a claude-code, and Cursor or another agent uses the interactive `npx skills add bitjaru/styleseed` and then Project and Copy. There is also a warning for anyone hunting through the interactive list, because Codex is already included under Universal and you should not look for it again under Additional agents.

The seventy four rules, and the checks that make them bite

StyleSeed ships 74 craft rules and a 0 to 100 scored gate, and the gate tells an installed agent to revise screens below 80 before presentation. The checks must be interpreted through the selected grammar and the project contract rather than as universal taste, and the README is blunt about the limits: the workflow must actually be invoked, a markdown file alone cannot block delivery, and 80 points is not expert approval.

That last set of statements is the honest framing, and it is what separates this from a rule file you paste into a system prompt. A markdown file that nobody reads is not a gate. The gate exists only when the workflow runs, and the README says so rather than implying the rules are self-enforcing.

The rule table gives a representative sample of what is checked and why. An unselected default accent requires an intentional, recorded color role. Emoji used as UI icons are rejected because they inject random colors and render differently per operating system. The icon-in-a-chip above every feature card is called a generation one AI tell, decoration pretending to be information. An all-even grid of same-weight centered cards is named the number one machine-composed tell, because there is no focal point. Accidental flat-black backgrounds should use the selected grammar's surface model. Type sizes outside the selected scale prevent unintentional hierarchy drift. And hardcoded hex in components are rejected, because tokens only, otherwise the lock cannot hold across screens.

That last one is the load bearing rule in a system whose whole purpose is consistency across screens. If a component hardcodes a colour, the palette is no longer the source of truth, and a rule set that cannot be enforced against the code is a preference list.

Skins, grammars and the difference from a token set

What the published demonstration shows is how far the system reaches beyond colours.

There is one component and three token skins, with the same chat UI morphing across Toss, Raycast and Arc inspired token sets. The README is precise about the division: a skin changes colour and typography, while a StyleSeed brand recipe changes the actual geometry, containment, controls, collections, density and motion. And the skins are described as inspired-by token sets, not recreations.

That distinction is the one to hold on to, because it is where most design system tooling stops. A token set is a palette and a type scale. A brand recipe is a set of structural decisions about how components are shaped, how dense they are, how they move. If you only want the colours, the skins are enough. If the problem is that every generated screen has the same flat, evenly spaced, centred layout, then the palette was never going to fix it and the recipes are the part that matters.

The word grammars appears throughout, and it is the mechanism for that. A grammar is a selected set of rules with a particular vocabulary and a particular set of exceptions, and the checks are interpreted through the grammar you selected rather than applied universally. The project also references semantic palettes, reference compilation and evidence-verified UI in its description, and the number of structural recipes is nine, with a page comparing them.

The examples/ directory holds two projects, an incident-workbench and a wanted-design-case, which suggests the intended usage is as a worked case rather than a generic utility.

Evaluating it, and the version pinning question

The README has a section for people assessing the repository, and it points at a three-minute verification path to test a clean public install before trusting the README. A source-level evaluator quickstart then connects that install to the published release manifest, Windows PowerShell evidence, CI, an SBOM, the benchmark boundary and the production build.

That list is a checklist of the things you would want to see before adopting a tool that rewrites how your agents work, and the fact that the project documents it is a good sign.

The versioning note is the part to decide on. For a fixed evaluation version the README points at the v5.0.0 release archive and its checksums, while the default bitjaru/styleseed command follows a mutable main. So there are two ways to install and they behave differently, and for an evaluation you want the first one.

There is also an upgrade path for anyone who already installed the 23-skill layout. You reinstall through the same channel and then ask StyleSeed to consolidate your old skill entries. The update workflow verifies the new unified payload and archives only unmodified old ss-* directories outside skill discovery, so modified copies remain for review. Then you restart the agent session, and the registrations become $styleseed build and /styleseed build instead of the former $ss-build and /ss-build. All 22 workflows remain bundled, and selecting only styleseed is now sufficient.

That migration design is worth noting. The archive rule means an agent skill someone has edited by hand does not get silently deleted, and the fact that the v5 release is named One skill, 22 workflows describes the consolidation from many entries to one.

Editorial conclusion

Adopt StyleSeed if your team is generating UI with coding agents and the output keeps converging on the same generic look, because the rules are compiled from your own recorded decisions and the score gate turns taste into a check the agent has to satisfy. Do not adopt it expecting expert-level output, since the README says the goal is repeatable application of expert judgment rather than expert-level output today, and a score is not a human approval. Before you install, read the product principles and the roadmap, and remember that the default install follows a mutable main branch, so use a tagged release with checksums if you need a fixed version.

Frequently asked questions

Official sources

1. bitjaru/styleseed on GitHub
2. License: MIT
3. Project website
4. README
5. Releases