Turns a plain-text request into a working interface foundation for websites AND web apps — assembled fast, machine-verified, testable locally, and ready for further UI refinement whenever you want. Not a pretty-page generator: a basic but real product that works from the first hour.
Open-source pipeline for production-ready UI — the core is fully free (MIT), forever: all skills, all gates, base skins, base packs, the D1–D38 floor, the knowledge vault. Ecosystem assets (curated packs, skins, starters) and support are the future Pro tier — a bonus, never a paywall on capability.
K0 discovery → K1 structure → GATE 1 (human) → K2A base skin (automatic)
→ K3 verification → [K2B full visual craft — optional, anytime] → [K4 deploy — optional, GATE 3]
A skill pack for Codex CLI-style agents. Give it "make a landing page", "prototype a dashboard with two roles", or "I have a prototype, make it look good" — and it runs the request through a production line: structure first (a neutral skeleton approved by a human), an automatic calibrated base skin, then machine-verified quality. Full visual craft is an option you can trigger today, next month, or never. The human makes the taste decisions; the machine proves the floor; AI assists only where it is reliable.
- The first artifact lands in minutes. Verified Starters (
starters/) ship pre-verified structure; copy injects through a slot map; finished projects feed new starters back through the harvest flywheel. - Real products are apps, not landings. The
appprofile models the domain, RBAC, user flows, a state matrix and an OpenAPI contract with typed mock fixtures before screens. - Taste work never blocks a working product. The base skin applies automatically after Gate 1; full visual work (K2B) is a restyle route over the finished foundation, never a rebuild.
- No AI slop. Divergence is constructed and machine-verified; a linted
ban-list blocks the statistical defaults, every ban citing its evidence
note in
knowledge/. - Quality is proven, not claimed. A deterministic floor D1–D38,
tiered (blocking / warning), including visual regression and a secrets
scan; a missing capability reports an honest
unavailableand caps the verdict.
| Skill | Conveyor | Owns |
|---|---|---|
pipeline-orchestrator |
— | routing (site/app, starter-first, neutralize, restyle), the contract as single source of truth, gates, packs, decision log, delivery report |
structure-builder |
K1 | UX — experience model, domain + RBAC + state matrix (app), neutral skeleton, neutralization |
visual-director |
K2A/K2B | UI — automatic base skin; directions, blind choice, merge, DTCG tokens, assets (K2B) |
quality-guardian |
K3 | proof — deterministic floor D1–D38 (tiered), AI diagnostics, quality report, verdict |
Sidecars: packs/ (integration bus: manifest + acceptance test + frozen
registry per integration — 15 packs in v7.0, incl. ai-look-detector,
contrast-checker, token-validator, state-generator, copy-linter,
awwwards-reference), starters/ (Verified Starters + harvest),
skins/ (base-site, base-app — three-layer tokens + dark theme),
knowledge/ (evidence vault, 33 notes — Obsidian is an optional viewer,
never a dependency), radar/ (weekly ecosystem digest), showcase/
(gallery auto-built from starters).
- Gate 1 (structure) — the only mandatory human gate of the core: approve a picture (sitemap / flow-map + RBAC), not YAML.
- Gate 2 (visuals) — deferrable; runs only when K2B is requested. Blind contact sheet, merge by default.
- Gate 3 (prod) — only with an active deploy pack; explicit confirmation and a passed dry-run rollback.
- Any gate can be explicitly delegated to the machine (
provisional_ai, recorded, with a confirmation offer on return).
Modes: quick / standard / full by project size and risk; interactive (default) or autonomous interaction.
bash install.sh <target-repo>Copies the package, checks dependencies (Python ≥3.10 + pyyaml, Bash
≥3.2-compatible; playwright optional — browser checks degrade honestly
without it), runs the self-test, prints the ready line. See INSTALL.md
for manual steps and the v5.x → v6.0 migration (contract-migrate.py).
First run: prompt P01 from eval/example-prompts.md, score with
eval/eval-rubric.md (pass ≥ 10/12).
bash tools/dops selftest # the tooling's own acceptance probes
bash eval/selftest/run-self-test.sh # the package suite
bash eval/e2e/rehearse.sh # the end-to-end rehearsal CI runsThe third one matters most before a pull request. eval/e2e/rehearse.sh is
the only definition of the end-to-end job: it materialises a project from a
starter, emits the self-service panel, serves the build, records the visual
baseline and runs the whole floor against it. CI does not repeat those steps —
it installs an environment and calls this script.
That is deliberate. The steps used to live inline in the workflow, so checking
a change locally meant retyping them, and a retyped copy that carried one file
more than the real job hid a blocking defect until it reached main. A
rehearsal has to be the thing it rehearses; a self-test probe fails the build
if the steps ever drift back into the workflow.
Without playwright the script says so and the browser lane reports
unavailable — a thinner run never passes for a full one.
- v7.2 — the machine contour: the deterministic floor collapses from ~25
script invocations into a single
tools/dops verify(one command, one JSON), because cost and wall-clock of an agent pipeline are linear in the number of turns, not in the size of the rule base. Resident rules drop from ~12.9k to ~1.1k tokens via an explicit taxonomy (generative/generative_and_verified/verified/process): a rule belongs in the prompt only if it changes what the model writes, and what a checker catches is left to the checker. Adds[E.3]reuse by input hash with targeted invalidation; invariants A.25 gate-overtaking, A.26 "nothing silently", A.27 external-service boundary; checkpoints and a control queue (У-2/У-3) so the owner can steer a run without stopping it; and the revision conveyor П-1/П-2/П-3 — always-on pins, a four-lane sorting station, and a checker at the door that refuses with alternatives before any work is spent. AC-23 gains its inheritance exemption: the quick-mode ceiling limits production, not what a Verified Starter carries in. Includes the v7.1 work that never got its own entry (wiki-sync with pinned sources, gate enforcement at the stage boundary,install.sh --update). Contract schema stays 7.0 — no migration is required. - v7.0 — verified design intelligence: invariants A.12–A.24 and evidence vault E.22–E.38 (33 notes) from the formalized UI/UX rule base (ux-wiki); deterministic floor extended to D1–D38 with tiered enforcement (WCAG contrast, touch targets, focus, semantic HTML, reduced-motion, AI-look — blocking; squint/grayscale/keyboard/copy — warning); three-layer token architecture (primitive → semantic → component, compile-enforced) with dark-elevation and motion tokens; K2B upgraded (design review audit, Three Dials calibration, curated reference directions with hard stops); AJTBD-lite + RAT-lite discovery; slash commands (/design, /restyle, /verify, /deploy); six new verification packs; ecosystem MVP (Radar weekly digest, Docs Portal, Showcase); monetization groundwork (trademark notice, pro-tier structure, install-time tier check) — the core pipeline stays fully open source.
- v6.0 — working interface foundation: K2 split (automatic base skin + optional deferrable visual craft), app profile (domain, RBAC, state matrix, OpenAPI mocks), Verified Starters with harvest flywheel, pack integration bus, floor D1–D24, knowledge vault, contract schema 6.0.
- v5.2 — macOS/BSD hardening, PyYAML contract reads, flake-free checks, D5/D15/D21 fixes, self-test circuit.
- v5.1 — canonical D1–D21 registry, ready-family verdicts, degradation matrix, execution economy, decision log.
Code and documentation: MIT — see LICENSE. Brand (the "Design-Ops"
name, logo, and associated marks): trademark of Pavel Evgrafov — use
requires explicit permission (see the trademark notice in LICENSE).