The category contains four different jobs
The label "AI design system generator" hides a source decision. UX Pilot says it can start from a description or import a website, Figma file, repository, codebase, uploaded files, or DESIGN.md. DesignMD and Design Extractor begin with a website. designmd.app begins with a catalog of prepared files. All four routes may end with something called DESIGN.md, but they don't establish the same authority.
| Generate | Import | Extract | Reuse | |
|---|---|---|---|---|
| Starting state | No approved system yet | An approved system already exists | A website is the available observation | A starting style is enough |
| Typical input | Brief or visual references | Figma, code, repository, files, or DESIGN.md | Public URL or domain | Prepared DESIGN.md from a library |
| What the route does | Originates design decisions | Translates existing decisions | Infers decisions from rendered evidence | Adopts a prewritten direction |
| Main review question | Are the generated decisions suitable? | Did translation preserve the approved source? | Which claims were observed and which were inferred? | Is the file current, licensed, and appropriate for this project? |
| Common mistake | Treating generation as approval | Letting the import replace source authority | Treating reconstruction as an official system | Treating a reusable style as project-specific truth |
Extraction is observation, not approval
A website extractor can recover useful colors, type choices, spacing patterns, and component clues. It can't establish that every observed value is canonical, current, reusable, or permitted for your project. Record the URL and capture date, then mark inferred rules as inferred.
Capability matrix observed November 19, 2026
The matrix includes only what the captured product pages substantiate. "Unknown" means the evidence packet didn't establish the capability. It doesn't mean the product lacks it. Product pages also don't support a fair ranking of output quality, reliability, generation time, or consumer compatibility.
| UX Pilot | DesignMD | Design Extractor | Context.dev | |
|---|---|---|---|---|
| Starting input | Prompt, URL, Figma, repository, codebase, DESIGN.md, or files | Website URL, with remix inputs also described | Website URL | Domain or website URL |
| Creation mode | Generate or import | Extract, then edit or remix | Extract | Screenshot and style-data extraction followed by model-written output |
| DESIGN.md evidence | Documented as the central editable file | Documented output with an on-page sample | Documented output for coding agents | Documented portable design.md output |
| Editability | Token edits update the file | Manual and AI editing documented | Unknown in the captured excerpt | Unknown in the captured excerpt |
| Revision evidence | Draft, publish, and version history documented | Undo and history documented | Unknown | Unknown |
| Companion artifacts | Component library plus Figma or HTML output | HTML preview and optional Figma material | Unknown in the captured excerpt | Underlying style, brand, screenshot, and crawl data available through related APIs |
| Delivery routes | Product editor, Figma transfer, and HTML download are documented | DESIGN.md output, HTML preview, and Figma add-on | File placement, browser extension, and MCP are described | Interactive generator and API access |
| Important unknown | No cross-tool compatibility test | No independent accuracy test | No independent extraction-accuracy test | No independent extraction or consumer test |
| Banani | designmd.app | Stitch | Figma plugin listing | |
|---|---|---|---|---|
| Route | Prompt or visual-reference generation | Prepared-file selection | Format import and export | Figma-based generator signal |
| Captured output evidence | Components with Figma and code exports | Downloadable prepared DESIGN.md files | Portable DESIGN.md import and export | Evidence-insufficient capture |
| DESIGN.md substantiated here | No | Yes | Yes | No usable body evidence |
| Decision use | Consider for editable components and code delivery, then verify whether DESIGN.md is required | Consider when a reusable starting direction is acceptable | Consider when format portability is central | Verify through stronger first-party evidence before relying on it |
This normalization avoids a common buying error. MCP, API access, download, and copy-paste describe how material travels. Figma material, component code, HTML, token files, and DESIGN.md describe what travels. A product can offer an effective delivery route for the wrong artifact.
Choose from the source state you actually have
No system exists yet
Use prompt-led generation when the project still needs design decisions. Before generating, write down the audience, product surfaces, modes, required states, constraints, and intended consumer. The output remains a proposal until someone with authority accepts it. A polished file doesn't make the choices final.
An approved Figma or code system exists
Prefer import or direct derivation from the approved source. Record the exact file, repository revision, package version, or existing DESIGN.md used. The acceptance question is preservation: did the generator carry the approved decisions into the new artifact without adding silent substitutions?
A public website is the only source
Use extraction when the observed implementation is genuinely the best available evidence. Capture the URL, date, pages inspected, viewport or mode where relevant, and anything the tool couldn't see. Separate measured values from interpreted rules. A card radius may be observable; the reason it applies to one component family and not another may not be.
You need a reusable starting style
A prepared-file library can be faster than recreating a direction from scratch. Treat the selected file as a candidate source. Check its revision, scope, dependencies, attribution or license terms, supported modes, and intended consumers. Replace generic brand claims with project-specific decisions before calling it approved.
Inspect a complete public kit before choosing a route
Use Ambient Sage to see DESIGN.md beside semantic tokens and separate implementation exports. Inspect the artifact boundaries first, then decide whether your project needs generation, adaptation, or direct application.
Inspect the actual DESIGN.md before comparing feature lists
Start with identity. You should be able to tell which source produced the file, when it was produced, which generator and settings were involved, and whether the file is a draft, published revision, extraction, import, or prepared template. Without that record, two files with the same name can silently represent different systems.
- 1
Establish source authority
Name the approved brief, Figma file, repository revision, website capture, or prepared file. State whether the source originates decisions or only supplies observations.
- 2
Inventory decision coverage
Check colors, typography, spacing, layout, modes, component treatments, states, motifs, assets, usage rules, and explicit prohibitions. Mark absent categories as absent. Don't fill them by assumption.
- 3
Separate observation from inference
For extracted or generated material, label which values were measured, which rules were interpreted, and which decisions were newly proposed.
- 4
Record unresolved decisions
Keep missing states, conflicting rules, unsupported assets, and consumer-specific questions visible. Unknown is a useful status.
- 5
Capture revision and regeneration evidence
Record the artifact identifier, generation date, generator version where available, settings, source revision, and trigger that requires regeneration.
Use the first unsupported claim as a stopping point
If a file says a component is accessible, production-ready, or compatible with a named tool, look for evidence that supports that exact claim. A valid Markdown file or successful download doesn't establish any of them.
Keep artifacts separate from delivery routes
| Artifact | Delivery route | |
|---|---|---|
| Examples | DESIGN.md, token file, component source, HTML, Figma material | Download, copy, registry, CLI, MCP, API |
| What it establishes | The form and content available to a consumer | How a consumer obtains or applies material |
| What it does not establish | Correct project mapping, adoption, rendering, or compatibility | Which exact revision arrived or whether it was used correctly |
| Acceptance evidence | Identity, revision, contents, dependencies, and precedence | Resolved artifact identity, destination, operation result, and repeatability |
DESIGN.md is usually strongest at intent, rationale, usage guidance, and constraints. Token files carry reusable values and references. Component or HTML code carries implementation structure, while Figma material supports design editing. None automatically replaces the others. If two artifacts express the same decision, define which one governs and how the other is regenerated.
The same discipline applies to MCP and APIs. Structured access can make discovery and retrieval convenient, but access isn't adoption. Record the artifact identity returned by the route, where it was written or mapped, and which consumer read it.
Copy the export-readiness record
Complete this record before approving a generator for broad use. It connects the source, artifact, delivery, and consumer checks that product landing pages usually discuss separately.
Export-readiness record
Decision state
- Record status: proposed | accepted | revised | blocked
- Recorded on:
- Owner:
Source
- Source type: brief | Figma | code | repository | website | prepared file
- Source identity:
- Source revision or capture date:
- Authority: approved | candidate | observed | inferred
- Scope and exclusions:
Generation
- Product and route:
- Generator version, if exposed:
- Settings or prompt identity:
- Generated on:
- Regeneration trigger:
DESIGN.md
- File path or immutable identity:
- Artifact revision or hash:
- Included decision categories:
- Observed facts:
- Inferred or newly generated decisions:
- Unresolved decisions:
- Edit and revision evidence:
Companion artifacts
- Token artifact and revision:
- Component, HTML, or Figma artifact and revision:
- Fonts, assets, and dependencies:
- Precedence when artifacts disagree:
- Synchronization or regeneration method:
Delivery
- Route: download | copy | registry | CLI | MCP | API
- Destination:
- Retrieved or written artifact identity:
- Operation evidence:
Named consumer
- Project and revision:
- Component or screen:
- Mode, state, viewport, and content case:
- Project mapping:
- Expected result written before application:
- Observed result:
- First divergent layer:
- Accepted exception, owner, and review date:
Decision
- Accept, revise, or block:
- Evidence still required:
- Next bounded test:A hash is identity evidence, not quality evidence
A file hash can prove that two consumers received the same bytes. It can't prove that the design decisions are suitable, complete, or correctly rendered.
Ambient Sage as a bounded artifact inspection
Ambient Sage is useful here because its public page exposes a specific, versioned artifact boundary. The page identifies version 1, updated July 30, 2026, and offers a DESIGN.md. It also documents Plus Jakarta Sans for heading and body roles, JetBrains Mono for technical strings, a compact product-oriented type scale, semantic light and dark tokens, and separate DTCG, CSS, Tailwind, and shadcn export families.
Token specimen · real values
Ambient Sage
Live renderAmbient Sage's actual tokens — the same values its exports use.
This establishes that the kit has a written brief and companion implementation artifacts. It doesn't establish that an arbitrary project maps those artifacts correctly, loads the fonts, supports every component state, or renders the intended result. Those claims need a named consumer test.
A bounded intake record could identify Ambient Sage v1 as the upstream source, list DESIGN.md plus the required export family, and leave the project mapping and observation fields open. That's more honest than marking the kit "implemented" because the files downloaded successfully.
Failure modes that should change the buying decision
- Extraction is mistaken for approval. The tool reconstructed a website, but nobody established which observations are canonical or reusable.
- A prepared file is stale. Its catalog entry remains available while dependencies, agent instructions, or the target project have changed.
- Generated prose hides unsupported inference. The file states component rules or design reasoning that the source never established.
- Companion artifacts drift. DESIGN.md describes one value or rule while tokens, Figma material, or component code carry another.
- The advertised route doesn't identify the artifact. MCP or API access works, but the team can't record which revision was retrieved.
- The consumer mapping is unsupported. The file exists in the repository, but the agent, build, component, or application doesn't read it.
- Local overrides win. A project stylesheet, wrapper, breakpoint rule, legacy component, or copied value replaces the intended system.
- The team edits upstream too early. A rendered mismatch appears, and the source file is changed before anyone checks delivery, mapping, component use, or local overrides.
These failures aren't equal. A stale source requires a source decision. A mismatched token export requires regeneration or transformation work. A local override belongs in the consuming application. Changing DESIGN.md for every visible mismatch can make the authoritative artifact less accurate while leaving the real fault untouched.
Run one representative-screen test
Don't approve a generator from its sample gallery alone. Pick one screen that exercises meaningful design decisions: primary and secondary actions, text hierarchy, a form or data component, at least one interaction state, and the relevant color mode. Keep the test narrow enough to inspect every handoff.
- 1
Write the expectation first
Name the source decision, artifact field, component or screen, state, mode, viewport, and visible result you expect. This prevents the rendered output from redefining success after the fact.
- 2
Confirm the exported identity
Record the DESIGN.md revision and every companion artifact used. Confirm that the delivery route retrieved or wrote those exact identities.
- 3
Inspect project mapping
Check that the project loads the intended files, resolves token names, supplies required fonts or assets, and exposes the guidance to the named agent or builder.
- 4
Inspect the consuming component
Confirm that the component supports the intended variant and state and actually consumes the mapped values or rules.
- 5
Observe the rendered case
Record what happened under the chosen mode, viewport, content, and state. A screenshot can support the observation, but retain the project and artifact revisions with it.
- 6
Correct the first divergence
Fix the earliest layer where evidence stops matching the expectation. Don't regenerate or rewrite the upstream artifact unless the divergence begins there.
What this test can prove
A passing representative case supports acceptance for that recorded source, artifact revision, project, component, state, mode, viewport, and environment. It doesn't certify every screen or establish a universal product ranking.
Make the purchase decision
Shortlist a generator only if its documented starting input matches your source state and its substantiated output matches the next consumer. Then inspect one real DESIGN.md, record its companion artifacts and unknowns, and run the representative-screen test. If the vendor can't expose enough identity, revision, or artifact detail to complete the record, treat that as a buying constraint instead of filling the gaps with assumptions.
For a project with no approved direction, start with generation and review the decisions. For an established system, favor preservation through import. For a website-only reference, extract but label inference. For a reusable starting point, choose a prepared file and adapt it deliberately. Next, complete the export-readiness record for one candidate and one screen before comparing prices or expanding the trial.
Sources
- UX Pilot AI Design Systems Generator: UX Pilot documents prompt-led creation, imports from several source types, an editable DESIGN.md, draft and publish states, version history, and separate Figma or HTML outputs.
- Banani AI Design System Generator: Banani documents prompt or visual-reference generation, component variants, Figma transfer, direct code export, and MCP integration, but the captured page does not establish DESIGN.md output.
- DesignMD: DesignMD documents website-to-DESIGN.md extraction, manual or AI editing, history, an HTML preview, and a separate Figma add-on.
- designmd.app: designmd.app presents a library of prepared DESIGN.md files intended to be selected, downloaded, placed in a project, and configured for a coding agent.
- Google Stitch DESIGN.md announcement: Google describes DESIGN.md import and export in Stitch and presents the format as a portable way to carry design rules and reasoning across platforms.
- Design Extractor: Design Extractor documents URL-based DESIGN.md extraction and handoff through file placement, a browser extension, or an MCP route.
- Context.dev Design.md Generator: Context.dev documents domain-based extraction using screenshots and style data, with related API access to the underlying data services.
- Ambient Sage Design Kit: The public Ambient Sage page identifies a versioned free kit with DESIGN.md, semantic light and dark tokens, typography inputs, and separate DTCG, CSS, Tailwind, and shadcn export families.
- How to generate a DESIGN.md: Identity Forge distinguishes the written DESIGN.md brief from the exact token values and implementation artifacts that accompany it.
- Design system generator with code export: Identity Forge distinguishes exported artifacts such as DESIGN.md, tokens, HTML, Figma material, and component source from delivery routes such as download, registry, CLI, and MCP.