Seenry skills
Agent skills that make an AI work like a design studio. Seenry fixes a strict frame (pixel grid, columns, padding, safe space, type set), explores several compositions inside it, critiques them side by side and refines one with real material, using measured evidence from real company screens through the optional Seenry MCP.
| Skill | Use it for |
|---|
| seenry | Building any web page or component: system, layout, type, color, craft, copy, accessibility, imagery, reference matching |
| seenry-assets | License-clear photos, icons and fonts from the web, generated images when an image model is available, provenance for every file |
| seenry-review | Auditing a screen, reviewing a UI diff, stress-testing one component, explaining how something is built |
| seenry-motion | Transitions, gestures, scroll and component animation, with runnable assets |
| seenry-apps | Native-feeling mobile screens and flows |
| seenry-video | Product films as code: launch videos and explainers studied from real references, rendered frame by frame, with designed sound and sourced claims |
What makes it different
- It looks before it designs. With Seenry MCP the agent searches a library of more than 100,000 captured app screens and nearly 2,000 websites, matched by category ("analytics hero", "payments pricing") as well as by name. It surveys up to eight real references side by side, keeps the three or four that fit the brief with a reason for each, measures their type, radius and spacing, then designs its own page and credits what it kept in
DESIGN.md.
- A system before styling. 4px grid, one spacing scale, one radius family with concentric nesting (
inner = outer − inset), one sans family (plus mono for data), one accent color, rings instead of heavy shadows. Starter tokens.css and a DESIGN.md template keep page 12 consistent with page 1.
- Studio process. Brief → research → frame → material → explore 3 compositions → critique → refine → verify. Structured exploration varies arrangement, scale, anchors and density inside the frame, never the frame itself.
- Real material.
seenry-assets sources CC0 and public-domain photos (Openverse), permissive icon sets (Lucide, Phosphor, Tabler via Iconify) and open fonts, or generates art-directed images through OpenAI or Gemini, and writes a license manifest.
- Components built in layers. Grid → safe space and content areas → structure → type and states, recorded as a short spec card. Max three sizes and three weights per component. Anatomies with measured defaults for buttons, inputs, cards, rows, menus, dialogs, tables, pricing tiers and more.
- Page shells and section archetypes so every page shares gutters, widths, rhythm and heading patterns. Pages.
- Measured, not imagined. Benchmarks from 18 leading sites (Linear, Stripe, Vercel, Notion, Figma, Raycast, Resend, GitHub and more): hero sizes, weights, tracking, radii, elevation. With Seenry MCP the agent reads available captured measurements and studies real sections, pages, app screens and recordings. .
Measured results
A blind benchmark on October 7, 2026 compared Seenry with Seenry MCP against two public design skills (Emil Kowalski, Jakub Krehel) and no skill. Each condition ran Codex (Sol 6.1) on the same prompts a typical user writes: 10 edits inside two existing apps and 8 new sites from a one-line request. Seven of the 18 tasks were held out: the skill was never tuned on them. Claude Opus judged screenshots blind, with labels shuffled per packet, 80 judgments per condition on a 10-point scale.
| Seenry + MCP | Emil | Jakub | No skill |
|---|
| All 18 tasks | 7.99 | 6.84 | 6.53 | 6.61 |
| Edits, tuned tasks (7) | 7.69 | 7.19 | 6.78 | 6.69 |
| New sites, tuned tasks (4) | 8.25 | 5.88 | 6.62 | 6.56 |
| Held-out tasks (3 edits, 4 new sites) | 8.21 | 6.93 | 6.14 | 6.54 |
Against Emil's skill, Seenry won 59 judgments, tied 15 and lost 6, and led 15 of 18 tasks. It lost a recipe screen and an empty state and tied settings. It is slower: about 17 minutes per new site against 6 to 7, and 6 minutes per edit against 3 to 4, because it researches references, explores directions and runs its own review. After that run, the review loop was sped up (parallel critic and motion judge, verdict reuse on unchanged boards). A timing rerun at commit e650259 took 11 to 13 minutes per new site (4 sites that took 14 to 19 before) and 2 to 3 minutes for two edits, and blind Opus judges preferred the new sites in 5 of 8 pairings against the earlier runs (7.75 against 7.38). Measured at commit ed2a30f. One model family judged the final round, and scores describe these prompts, not every project.
Install
npx skills add ayangabryl/seenry-skills
Or clone and install locally (previews by default):
python3 scripts/install.py --apply
Replacing an older Seenry install: run with --replace --apply. Retired skill folders are archived, and the printed manifest works with --rollback.
Seenry MCP
The read-only Seenry MCP at https://mcp.seenry.design serves captured websites and sections, measured CSS evidence, recordings, and iOS app screens and flows. It needs a Seenry Pro key sent as Authorization: Bearer <key>; the server does not offer OAuth sign-in. .mcp.json reads the key from the SEENRY_PRO_KEY environment variable. Without MCP the skills fall back to the bundled benchmarks and ordinary browsing.
All eight skills use the Seenry MCP 0.1.1 tool contract and prefer connected MCP tools. Scripted research uses the same endpoint and an existing SEENRY_PRO_KEY: run node skills/seenry/scripts/research.mjs --brief BRIEF.md --family websites --terms "<focused query>" from this repository, or the installed skill's absolute script path. The script makes at most three tool calls (planned maximum 11 credits), saves the source context and images it used, and stops on service, auth or quota errors. Without a key it writes an explicitly unresearched handoff for connected tools or permitted web research.
Verify and contribute
python3 scripts/validate.py
python3 -m unittest discover -s tests
node tests/system-audit.browser.mjs --playwright /path/to/node_modules/playwright/index.mjs
The checks cover package structure and the included tools. They do not certify visual quality: render the page at 1440 and 390, run the system audit, and look.
Seenry is licensed under the Apache License 2.0; see NOTICE. Contributions follow CONTRIBUTING.md, including the Conventional Commits format. Adapted guidance retains its authors' notices in skill licenses. External references remain research sources; use their code and media only under their own terms.
