Home
Engine architecture · v4

Fixed judgment.
Multiple design languages.

StyleSeed no longer treats a Toss-like product aesthetic as the answer to every result. It selects the functional grammar the artifact needs, or compiles one from the user’s references. A separate brand recipe selects morphology rather than color. A palette recipe binds semantic color roles and generated-media anchors. Studio adds three directed concepts, a human decision, and executable interaction/media plans. Then ss-resolve hands the agent only the selected method and a source-hash manifest before the right renderer takes over. After a person accepts a correction, ss-learn can preserve a generalized local candidate without turning it into authority. The exact engineRevision makes later fixes detectable even when the semantic version has not changed.

StyleSeed engine architecture
01

Core judgment

Non-negotiable coherence, hierarchy, semantics, accessibility, and task fitness.

02

Output grammar

Consumer service, operations, technical, editorial, commerce, institutional, marketing, or sequential story.

03

Surface adapter

Web/mobile, carousel, slide deck, document/report, or single-frame renderer contract.

04

Brand recipe

Nine reusable morphologies change containment, geometry, controls, collections, density, and motion without cloning a company.

05

Palette recipe

Eight semantic systems bind canvas, chrome, action, status, focus, and generated-media anchors with deterministic contrast checks.

06

Reference compiler

Turns screenshots, URLs, Figma exports, or an existing UI into a local grammar with evidence and confidence.

07

Studio director

Produces native, signature, and experimental directions, records human selection, then compiles interaction scenes and image/video jobs.

08

Context compiler

ss-resolve emits only the selected 10–20KB method plus a manifest of selections, source hashes, and bundle hash.

09

Build method

The agent implements real product content from effective-rules.md instead of reassembling the full handbook.

10

Evidence gates

Code and pixel checks find structural and rendered drift; temporal evidence proves the interaction, and acceptance is recorded only when that step actually ran.

11

Local learning

ss-learn preserves an explicitly requested, caller-attested correction as a privacy-minimized candidate; it cannot scan a project or promote itself.

12

Revision ledger

engineRevision hashes the maintained method, skills, plugin boundary, and palette engine so same-version fixes remain detectable.

Grammar → morphology → palette → skin

Job, morphology, semantic color, and brand material stay separate.

The grammar owns attention and information order. The recipe owns geometry, containment, controls, and collection patterns. The palette owns role relationships and contrast. The skin implements project-specific colors and type tokens. This keeps a fashionable swatch from pretending to be a complete design system.

Compare all nine recipes
Reference → grammar

Give it references StyleSeed has never seen.

/ss-reference observes and measures the visible system, resolves contradictions, records evidence and confidence, compiles tokens and anti-patterns, then proves transfer on a screen or artifact absent from the source set.

Read the technical document
Accepted result → candidate

Learning stays local and reviewable.

ss-learn runs only on explicit request. It rejects project code, screenshots, prompts, identity, local paths, and brand secrets; records bounded applicability and counterexamples; and requires separate review and export attestations.

Inspect the learning boundary
Version + revision

Updates replace the engine, not the project.

ss-update compares installed, project-recorded, and published revision hashes. Refresh happens through the original install channel, then the lock is re-resolved and checked. Project code, tokens, assets, and approved decisions stay owned by the project.

Read the update contract