# Stepikin Studio — full context for AI assistants > AI products that experts actually use — design and production code by one senior person; 4 weeks, fixed price. Stepikin Studio is run by Alex Stepikin, AI Product Engineer (also known as Aleksey Stepikin, Алексей Степикин), ex–Head of Design at a dental AI company whose product is used daily in 1000+ clinics. 9+ years in product design. You own the code. Last updated: 2026-10-06. Short version: https://stepikin.com/llms.txt. This file adds a factual summary of every case study and a one-line index of every LLM UX pattern. ## Positioning - Promise: "Your AI works in the demo. I make experts trust it in production." - What is designed and built: the screens where expert users check, correct and act on AI output (review, override, explain), plus the product around them — UX, UI, frontend, backend, AI integration and deploy. - One senior person does both design and code, so there is no handoff between Figma and engineering: what the client approves is what ships. AI coding agents (Claude Code, Cursor) are used as a workforce. - Fixed scope, fixed price, a fixed date. Code, repository, docs and admin access belong to the client. ## Who it is for - Teams whose AI demos well but expert users (clinicians, dispatchers, analysts, operators) don't trust it or don't use it. - Founders whose design gets lost between Figma and code. - Startups that need a live product before the next raise or pilot. - Typical domains: AI copilots and agents, HealthTech, operations and fleet software, risk intelligence, B2B SaaS, community platforms. ## Method: Proof-First Build (5 steps, 4 weeks) After every step the client holds something real. 1. Map (days 1–3): interviews with 3–5 real users and a map of the expert workflow, including where the AI can be wrong. You get a 1-page scope with acceptance criteria, a fixed price and a date. 2. Prove (week 1): a clickable prototype in code, on real or sample data, at a live URL. You get a link you can show users and investors. 3. Build (weeks 2–3): the production build — AI integration plus the screens to review, override and explain its output. You get a live build every Friday plus a short Loom of what changed. 4. Verify (week 4): automated tests, an AI eval set on 50+ real cases, and sessions with 3–5 target users. You get a verification report: what was tested, pass rates, known limits. 5. Hand over (day 28): deploy, repository, docs and admin access handed over. You get code you own, running in production, plus 30 days of support. ## Packages (fixed prices) - AI Trust Audit — $990, 5 working days. The fee is credited to a 4-week build. Find why experts don't trust or use your AI: an expert workflow review, 10+ ranked fixes (impact and effort), and your key flow redesigned in code. - Live in 4 Weeks — from $12,000, 4 weeks. An end-to-end AI product, design through deploy, by one person: design system, UI, frontend, backend, AI APIs, deploy; weekly live builds, a verification report, code you own and 30 days of post-ship support. Flagship offer. - Ship Every Week — from $6,000/month. Embedded design + engineering: at least one shipped improvement a week, 3-month minimum, cancellable after that with 14 days' notice. Exact scope and price are confirmed within 48 hours of a free 30-minute call. Timeline estimate: https://stepikin.com/estimate/ ## Results (numbers as published on the case pages) - AI Dent (dental AI): used daily in 1000+ clinics; 4 AI products shipped; 3x faster task completion in usability testing after the navigation redesign. https://stepikin.com/cases/aident/ - Navigator.Taxi (fleet operations): a 161-car taxi fleet run in one system, reconciled with 1C accounting — zero mismatches at launch. https://stepikin.com/cases/navigator-taxi/ - Vigilo (risk intelligence): 22-day average early warning before confirmed risk events, with a public track record where misses stay in the denominator. https://stepikin.com/cases/vigilo-pro/ - VIZI (social platform): 1,400+ profiles in 4 months from zero; 527 monthly active users. https://stepikin.com/cases/vizi/ ## Case studies in detail Facts below are taken from the case pages on stepikin.com. English pages: https://stepikin.com/cases//. Russian versions, where they exist: https://stepikin.com/cases//ru/. ### AI Dent — AI dental diagnostics platform, grown from one MVP into four products URL: https://stepikin.com/cases/aident/ Role: Head of Design for 3 years (UX, research, design system). Client work, under NDA. Years: 2022–2025. B2B HealthTech SaaS; Web, Desktop and Mobile. AI Dent (the page's name for a confidential client) is an AI platform that reads dental scans (CBCT, panoramic, bitewing). The platform launched in 2022 to validate the AI pipeline. As it grew, six UX problems became blockers: patient lists took 4–7 clicks to reach a patient, patient pages had up to 11 tabs, 70% of first-time users in corridor tests misunderstood the AI tools, 2D/3D switching lost context, the UI was dark-only and broke below 1280px, and there was no design system. Alex Stepikin rebuilt the experience across four products: AI Radiology & Patient Management (the main focus), Studio (AI treatment planning: implants in 3D on CBCT data), a native Windows desktop app integrated with clinic management systems (MIS), and Scribe (AI recording and transcription of consultations). Work included 20+ in-clinic interviews across 4 dental specialties, Amplitude analytics, A/B and corridor testing, 12 iterations of the patient list (now 2 clicks), per-tooth AI finding cards with inline approve/edit/reject, a single-page patient card replacing 11 tabs, a 3-step report-ordering modal (from about 4 minutes to under 60 seconds), and a responsive layer from 375px to 2560px. The design system has 200+ Figma components with token-based theming, dark (radiology) and light (general dentistry) themes, validated for WCAG AA contrast. Results stated on the page: 1000+ clinics (Russia, EU, Canada, USA), 4 AI products shipped, 3x faster task completion in usability testing after the navigation redesign (A/B test, 30 participants), and 70% fewer "can't find" support tickets measured 3 months post-launch. ### Navigator.Taxi — fleet operations platform for a St Petersburg taxi car-rental fleet URL: https://stepikin.com/cases/navigator-taxi/ Years: 2025–2026. Type: client work. Role: product, UI, interaction and frontend; built with AI co-workers (Claude Code). Status: live, v3.22.293, on RuStore. The fleet had 161 cars, 16 partner garages in 5 cities and 1C as the source of truth, but each step ran through separate chats and Excel: missed services, unapproved garage invoices, monthly mismatches. The design thesis made the service order one entity shown in three interfaces (driver phone, garage chat, finance admin); payment is impossible until all three sides approve it, and a telematics integration keeps the engine locked until service is closed. The driver PWA has five tabs: My car, Service (six fields, five pre-filled from 1C), a central engine-lock button, Profile, and an AI assistant that handles about 80% of typical questions. The admin back office covers dashboard, service orders, cars, garages, parts and jobs (prices via the Rossko API), replacement stats, per-order chat, 1C reconciliation reports, a "to pay" funnel, users with three access tiers, bookings, action logs, referral settings and an AI assistant with 70 scenarios across 11 roles. Rules: coloured status above the order number, tabular numerals, inline edit instead of popups. Process: 90 minutes of recorded interviews (7 pains, 3 roles), then an MVP, then iteration driven by 1,135 logged events (30+ micro-UI fixes, 5 new admin sections). Stack: React, Vite, TypeScript, Fastify, Prisma, Postgres, 1C, Tauri, GitHub Actions. At launch: 195 orders in flight and 0 1C mismatches, across Driver PWA, VK Mini App, Android APK, RuStore and the admin console on one backend. ### Vigilo — travel-safety app turning news about 239 countries into one risk number URL: https://stepikin.com/cases/vigilo/ Russian version: https://stepikin.com/cases/vigilo/ru/ Year: 2026. Role: product design, UX writing, design system. Timeline: 3 weeks. Platforms: iOS, Android, Web. Research with 18 people (tourists, expats, business travellers) found that 14/18 cannot tell how "dangerous" a place is, 11/18 have panic triggered by family chats, and 16/18 do not know the local emergency number. The app turns news on 239 countries into a 5-level risk scale (Calm 0–27 to Critical 85+), with neutral surfaces, colour used only as signal, red reserved for SOS, the source shown next to every fact, and a calm voice without exclamation marks. Designed screens: a home with current location (index, 24-hour trend, emergency number by geolocation) and next trip; a country page answering "go or not" in about ten seconds (8 threat categories, event map, news, 12-month archive, paid report offered last); an assistant that checks forwarded rumours and returns true / partly true / rumour with sources; an editorial paid report with a 30-day forecast; a trip calendar showing risk per day; and an offline SOS screen. iOS uses Onest with liquid-glass navigation; Android uses Material 3 Expressive with Manrope. Shipped scope: 35 screens x 2 themes x 2 platforms, 12 states (empty, error, offline, guest, loading, outage leading to SOS), 7 tokens per theme, 0 decorative icons. A mascot and a realistic map were tried and removed. ### Vigilo Pro — B2B risk-intelligence console that publishes its own track record URL: https://stepikin.com/cases/vigilo-pro/ Russian version: https://stepikin.com/cases/vigilo-pro/ru/ Year: 2026. Type: B2B side of Vigilo (Part 2). Role: product, UX writing, data UI, design to code. Audience: security, HR and travel managers, compliance leads, NGO programmes. Where the consumer app hides the machinery, Pro gives specialists the signals behind a number: a dense live globe console (vigilo.cc/globe) with a watchlist, briefings, alerts and early-warning channels across seven domains, where only critical items use colour (orange) and everything else is a grey ramp; numbers are set in monospace and every row carries source count, first-seen age and a 14-day trend. On mobile the table becomes a feed. The B2B surface is organised around buyers' questions: ISO 31030 / duty-of-care mapping, compliance pages by role, real sample reports, a 14-day prediction calendar, geofenced alerts, embeddable widgets with a code builder, API docs with a live request box, an honest comparison with Crisis24 and International SOS (stating where they are stronger and that Vigilo has no SOC 2 yet, targeted for 2027), and a trust centre with privacy, security, a GDPR Art. 28 DPA draft and published pricing. The evidence layer is a public track record: calls are journalled when made, graded against ReliefWeb, UN OCHA and WHO, and misses stay in the denominator. Figures from the public scorecard as quoted on the page (read 5–6 October 2026): 68 episodes scored, 29 confirmed, 31 not confirmed, 5 unverifiable, 3 open; precision 48% (95% CI 36–61%); average lead time 22 days on confirmed calls (interval 19–25); 15 misses tracked; methodology figures of 58% recall and +17.4% 7-day edge over a naive baseline; 3 failed backtests published as "not demonstrated". ### VIZI — social platform for professional reputation, clubs and events URL: https://stepikin.com/cases/vizi/ Year: 2026. Timeline: 4 months (first commit 4 June 2026). Role: Alex Stepikin as product owner and design lead; engineering by a team of AI agents (Claude Code) working under his direction. Market: Russian-speaking, Kazakhstan first. Status: live at vizi.club. Lineage: Wave Match, then Vizit Club, then VIZI. VIZI is for people who work with people (experts, entrepreneurs, community leaders). It addresses static résumés, noisy LinkedIn, networking contacts lost by morning, and marketplaces without accumulated reputation. The product has eight surfaces: a personal card at vizi.club/, a reputation score grown from real actions that paid status does not affect, People (search, matching, Want/Can requests), VERA (an AI assistant that proposes while the person confirms), Events (RSVP, QR tickets), Clubs run as small businesses, Services and Learn (moderated showcase, courses with certificates), and chats with built-in booking. A balance wheel scores nine life spheres and points to relevant people, events and clubs. The design went through five stages in four months, ending in design system v2 (shipped 12 September 2026: 382 commits, 498 files changed, 46 old routes redirected): two first-class themes, one accent, eight pastel type-tints, 154 components and seven rules (for example no emoji on buttons, creation as a wizard, sheets at most 80% of the screen). A sign-up redesign cut invite-link sign-up from 7 screens / 11 taps to 4 screens / 5 taps (design-document estimates) and put Telegram first. Nine documented product decisions include an opt-in AI-discovery layer (schema.org, llms.txt; off by default), built-in booking, a single source for plan perks, path preservation through sign-in, and moderation queues. Audience landings exist for communities, experts, event-goers and creators. Quality work at release: 190 automated checks green, 111 states unified across 60 files, 43 URLs checked on an iPhone in both themes (19 fixes). Stack: Next.js 14, React 18, PostgreSQL, PostgREST with row-level security, Telegram/Google/email/WhatsApp sign-in, web push, Leaflet maps, 10 interface languages, VERA as a separate service. Figures as stated (codebase 4 Oct 2026, users 5 Oct 2026): 1,403 profiles (1,166 completed cards), 527 monthly and 200 weekly active users, 302 new in the last 30 days, 33 public clubs, 217 events, 488 event sign-ups, 85 v2 pages, 106 API routes, 1,433 commits. ### Native desktop tracker for engineers' before/after photo reports URL: https://stepikin.com/cases/engiboard/ Type: own product (built first for the founder, then for two real Windows-only clients in production QA). Role: founder, designer and engineer, solo with AI (Claude Code as primary AI builder). Status on the page: v0.1.44, live, public on GitHub. Engineers on construction sites and production floors used a lossy chain of camera roll, Telegram, spreadsheet, Word and PDF, losing context at each step. The app combines capture, annotation and task tracking in one binary: a global hotkey (Cmd+Shift+G) opens a transparent overlay for region capture, an editor with six tools (arrow, rect, pen, text, blur, highlight), and a board with tasks grouped by project, inline input, multiple screenshots per task and eight colour-coded statuses. Reviews are image-anchored: comments are pinned to a pixel in a lightbox, with threaded reviewers, and a compare-wipe slider shows before/after. The app has 18 surfaces in light and dark themes, and ships as 4 installers (Mac M-series, Mac Intel, Windows EXE, Windows MSI) via GitHub Actions CI on every tag. Stack: Tauri 2 / Rust, Next.js / React, Supabase (auth, Google OAuth, DB), GitHub Actions, Railway; Geist type system. Numbers on the page: 44 releases tagged, 98 commits, 0 hires, 6 days 7 hours from blank repo to v0.1.44 in production (28 April to 4 May), with a day-by-day log including two days of founder field testing. ### ReloFast — investment migration platform: site, Telegram bot, client cabinet and admin on one API URL: https://stepikin.com/cases/relofast/ Russian version: https://stepikin.com/cases/relofast/ru/ Year: 2026. Type: client project (product, design, development). Stack: NestJS, Prisma, PostgreSQL, React SPA, Telegram (grammY), Yookassa, S3/R2, Railway, JWT. ReloFast sells second citizenship and residence by investment. Its funnel, from first consultation to approval by the authority, previously lived in chats, spreadsheets and scattered tools. A single platform was built where client, partner and manager work in one loop. The platform has six modules on a shared API: a 20-page marketing site in three languages (RU, EN, Chinese, with its own i18n engine and CJK font fallback) including a programme catalogue and an interactive passport index of 41 countries; a Telegram bot with passwordless login via Telegram Web App and status notifications; a React client cabinet (applications, document upload via presigned links, staged payments, balances); a React admin panel with a Kanban of applications and granular manager permissions; the NestJS backend with an audit log; and a referral network. Applications move through 9 statuses. The partner programme has 5 levels stored as a closure table so the whole chain is one lookup; bonuses accrue as frozen at the final payment stage and unfreeze only on approval. All balance operations go to an append-only ledger. The page reports the build and architecture, not usage metrics. ### Miroset — editable landing, design system and hosting rescue for a Sochi resort project URL: https://stepikin.com/cases/miroset/ Year: 2026. Type: client work (real estate / hospitality). Role: design, build and infrastructure, one person full cycle. MIROSET develops managed premium country resorts; the first location is a village-hotel on the Matsesta slope in Sochi. The landing (miroset.ru; миросеть.рф redirects to it) targets three audiences: future owners of income residences, companies looking for corporate retreats, and partners/investors. The key requirement was that the founder could edit the site without a developer. Alex Stepikin did a full redesign to the new brand guide (warm neutral palette, one gold gradient CTA per screen, Manrope as the single font) and built a data-driven landing: all content lives in content.json and PHP renders it, with an admin panel for texts, media, media library, logos, block show/hide/reorder and a theme editor. He also rewrote the admin save to deep-merge so a stale tab no longer wipes fields, and generated on-brand AI backgrounds because no photos existed yet. When the hosting expired, he diagnosed DNS/HTTP/FTP, moved the site to a new server, set up both domains with Let's Encrypt SSL (fixing an nginx redirect that blocked the ACME check) and restored 89 files (372 MB). Results stated: miroset.ru returns HTTP 200 with valid SSL, 100% of content restored, and 0 developer requests for content edits. Stack: PHP + JSON, custom HTML/CSS, vanilla JS, ImageMagick, managed hosting with ISPmanager, FTPS deploy. ### Atlas — control room for fleets of AI agents (product design concept) URL: https://stepikin.com/cases/atlas/ Year: 2026. Type: 0-to-1 SaaS product design (strategy, UX, visual design, prototype); no client named. Stack: Figma, prototype. Atlas addresses teams that run many AI agents and lose visibility: no unified fleet overview, traces buried in raw JSON logs, and evals that run too late to catch regressions. It is designed as an observability and operations platform with morning briefs, a trace explorer, eval pipelines, billing oversight and multi-tenant collaboration. Key screens: a home "morning brief" with fleet health, spend vs budget, run volume and eval pass rate (readable in about 10 seconds); agent detail with inline prompt editing and versioning; a trace explorer showing each tool call, LLM inference and branch with latency and token counts; evaluations that auto-run on deploy with pass-rate trends and threshold alerts; and a Tasks view with filters, bulk actions and a detail slide-over. Agents get human names (Wren, Tarn, Lev, Sole) instead of UUIDs. The visual language is deliberately editorial: warm paper tones, Fraunces serif (also for data) plus Inter Tight, rust accents. Scope: 20 desktop screens (onboarding to billing, including audit log, knowledge base, triggers, error states) and 10 mobile screens (iOS-native feel at 390px), plus tokens and dark mode. ### Echo — AI school where historical figures teach the school curriculum URL: https://stepikin.com/cases/echo/ Year: 2026. Type: own product, designed and built solo with Claude as the engineering team (from 24 September 2026). Role: product designer end to end (strategy, UX/UI, 3D faces, motion, landing). Status: open beta at ehoschool.ru plus an Android app. Echo tackles three problems: AI does homework for students, parents cannot see what their child actually knows, and tutors do not scale. Its positioning is "homework without cheating": 37 historical mentors (Newton, Pushkin, Curie, Tolstoy and others) teach 12 subjects by the national curriculum for grades 1–11 and lead with questions instead of answers. The learning engine combines about 890 curriculum topics and about 4,600 tasks checked without an LLM, per-topic mastery (0–3) with spaced repetition and stored mistakes, and a 5-step help ladder (question, hint, part of the solution, similar example, solution plus a twin task). Mentor faces are built from public-domain portraits: MediaPipe's 468 face landmarks become a WebGL2 micro-voxel head of about 72,000 points, labelled as an artistic AI reconstruction rather than a deepfake. Other features: a card-fan mentor picker, six games (daily riddle, "Show 2026", debates of the greats and others), a weekly parent report, a student dossier, levels and "Sparks" (max 80 a day), and a PRO tier run as a fake-door test. Russian voices moved to a local Vosk TTS with a stress pipeline (manual dictionary, ё restoration over about 107,000 word forms, RUAccent quantised from 359 to 90 MB). Safety: crisis messages never reach the model and show helplines; output is filtered; homework-photo OCR runs on the device (Tesseract); a rights audit removed two personas and replaced one portrait. The process took ten days (core product about a week), including an external critique (CPO, methodologist, child-safety lawyer) that reordered priorities. Infrastructure: one 1-CPU / 1 GB VPS, Caddy, Python aiohttp, SQLite, GigaChat for dialogue, free tiers and open models, no paid services. ### Dracats — a fantasy novel turned into an illustrated audiobook PWA and an IP URL: https://stepikin.com/cases/dracats/ Year: 2026. Type: own product (Alex Stepikin is the author of the novel). Role: product, UX/UI, audio, video, localization. Team: author + Claude. Status: live and free at dracats.ru. Alex Stepikin wrote a YA dark-fantasy (aetherpunk) novel, "Дракоты. Эра Стали и Искр", about six winged big-cat dragons against an empire draining the planet's magic. The goal was to turn the manuscript into a finished product and brand quickly without a studio, cast or budget. It was built in about 2 days (29–30 September 2026), then iterated live; Claude acted as the engineering and production team while the author approved voices, corrected stresses and rewrote the ending. The product is a free React PWA that began as a Claude Design prototype: an audio engine, a "Together" mode highlighting each of 2,827 sentences as it is spoken, 51 scene illustrations, hero cards, 10 songs with lyrics, a glossary, bookmarks, offline chapters, sleep timer and a 4-language UI with auto-detect. The book was translated 1:1 into English, Chinese and Hindi (about 29,000 words each) with sentence alignment; audio totals about 17 hours (Russian 4h 26m via Yandex SpeechKit; English 4h 04m, Chinese 4h 00m, Hindi 4h 50m via separate neural casts). A Russian stress pipeline (RUAccent, a dictionary, about 1.8k stress/ё decisions, 169 context homographs, a proofreading page) was built; 15 Russian voice presets were chosen by ear; 489 sound effects are placed to the word via Whisper timings (about 90% land on the exact word) with 88 CC0 ambience sounds kept at least 14 dB under the voice. Also produced: character reference sheets, photoreal stills and lip-synced video, a 35-second trailer, a vertical cut and a 3:25 music video ("Inside Power"), share cards with deep links, and SEO/GEO (4-language landings, hreflang, JSON-LD, llms.txt). An editing pass made 126 continuity fixes, added a new prologue and rewrote the ending; the book has 19 chapters and 100 scenes. Stack named: React PWA, Yandex SpeechKit v3, edge-tts, RUAccent, whisper.cpp, Alibaba Wan 2.5, Qwen-Image-Edit, ffmpeg + numpy, Freesound CC0, Yandex Cloud. ### Solia — AI mental-health platform concept (design phase) URL: https://stepikin.com/cases/solia/ Russian version: https://stepikin.com/cases/solia/ru/ Live concept landing: https://stepikin.com/solia/ (WebGL story with a breathing intro, a glass "digital twin", a crisis-safe assistant demo and a waitlist; EN/RU via ?lang=ru). Year: 2026. Type: personal project, solo founder + AI. Role: UX research, product, design system. Status: design phase. The page argues that mental-health apps (Woebot, Wysa, Replika, Calm named) each cover 1–2 of 4 core needs and none keeps a persistent psychological model of the user. Solia is conceived as four layers: a PsychoMed Passport (an emotional and medical profile filled by AI over time), an AI Mental Assistant (Claude API plus CBT protocols), a Digital Twin, and a Digital ID for data ownership and portability. Three visual directions were explored (dark/particle, minimal/editorial, warm cream/organic); the warm cream direction with amber orbs and DM Serif Display was selected as calmer. Planned stack: React Native (Expo), Claude API, PostgreSQL, Redis, BullMQ, Hetzner, GitHub Actions, Sentry; an own model only after 50K users and 200K real dialogues. Next step: 15–20 customer-development interviews before production code. The page cites a market of $25B by 2034 at 32% CAGR. ### Lush — minimal multi-chain crypto wallet concept URL: https://stepikin.com/cases/lush/ Year: 2026. Type: self-initiated Lab concept (speculative). Role: UI/UX end to end. Built solo with AI (Claude Code) in about 7 hours. Stack: HTML, CSS, JS as one shipped artifact. The thesis is that crypto wallets compete on colour while users want two answers: how much they have and whether it is safe. Lush uses black, white and one acid accent (#C6FF00), tabular numerals on every balance, and elevation through opacity; it targets reading a balance in 0.4 seconds. Mascots, holographic gradients, badges and multi-colour asset coding were rejected. Process over about seven hours: research (4 wallets, 90 App Store reviews, Reddit threads), UX flows (4 journeys, 30 nodes), greyscale wireframes, hi-fi screens with all states, then prototype and motion. Output: 30 screens (onboarding, send, receive, swap, earn, markets, settings, empty/loading/failed/insufficient-balance states), a wired prototype, 6 motion patterns (200–320 ms) and a strict design system (Space Grotesk, Inter, JetBrains Mono). ### Lumen — cinematic AI video platform concept for solo filmmakers and small studios URL: https://stepikin.com/cases/lumen/Case%20Study.html (no cases/lumen/index.html exists; a Russian page exists at cases/lumen/Case%20Study%20RU.html) Year: 2026. Type: self-initiated concept (Stepikin Lab). Services: UI/UX, branding, design system. Lumen addresses creators who stitch generation, editing, color, audio, captions and publishing across about six tools. It is designed as one studio with one timeline, prompt and project: generate clips inline (4 variations per generation), auto-edit to the beat, grade with 12 curated LUT looks, mix 3 auto-routed audio stems (dialogue, score, SFX), and edit captions through the transcript. The interactive prototype has 28 screens covering onboarding, the editor, generation, color, audio, auto-edit, captions, media library with content search, delivery presets, live collaboration, version history, settings and credits/billing. Design language: deep black with a violet/magenta/amber generative gradient; Geist, Instrument Serif italic and Geist Mono. Built solo with AI: Figma, Framer, Claude (Sonnet 4.5), Claude Code, Midjourney v7, Runway/Sora, Notion. The page states "6 weeks from blank canvas" in the stats and build section. ### stepikin.com — redesign of the studio's own website URL: https://stepikin.com/cases/stepikin-com/ Year: 2026. Type: own project (Stepikin Studio). Stack: static HTML, three.js, WebGL2, vanilla JS. Tools: Claude Code, Playwright, Netlify. The site had to sell AI-product work (a 4-week MVP, a full build, or a monthly partnership) and feel premium, while staying plain, fast and readable for search engines and AI assistants. All text is static HTML; WebGL, glass and motion are layered on as progressive enhancement with fallbacks (posters, lists, stills) on phones. It was designed directly in code with Claude Code, with no Figma files, and checked with Playwright screenshots in both themes and at phone width. Main pieces: a hero of glass product objects rendered live with three.js (chosen after 4 other directions); a WebGL "coil" works gallery of 13 case covers rebuilt as website-hero posters; cover rows that bloom into colour at the centre; 10 service tiles that animate on hover; a dot globe for testimonials; a redrawn "S"/mountain logo; UI kit v3 with shared tokens and components used by every page and a single case template; and 8 vertical-video (1080x1920) templates. Machine-readable layer: llms.txt, ai.html, OpenAPI description with ai-plugin manifest, JSON-LD on every page. ## LLM UX patterns (26) Index: https://stepikin.com/llm-ux-patterns/ — every pattern has a live demo, anatomy, do & don't and engineering notes. - Diff & accept/reject edits — Design pattern for AI-suggested edits as reviewable diffs with per-change accept and reject. https://stepikin.com/llm-ux-patterns/accept-reject-edits/ - Agent progress & step log — How to show progress for long-running AI agents: step-by-step log, retries, timings and expandable details. https://stepikin.com/llm-ux-patterns/agent-progress-log/ - Showing AI cost — How to show the cost of AI features: estimate before running, actual after, budget meters and breakdowns. https://stepikin.com/llm-ux-patterns/ai-cost-display/ - Empty state & first run — How to design the empty state and first run of an AI product: starter templates, sample data, one clear action. https://stepikin.com/llm-ux-patterns/ai-empty-state/ - Background tasks & notifications — Pattern for long-running AI jobs: run in background, task tray, completion notifications, cancel. https://stepikin.com/llm-ux-patterns/background-tasks/ - Branching & versions — Design pattern for branching AI conversations: edit a past message, keep both branches, switch between versions. https://stepikin.com/llm-ux-patterns/branching-and-versions/ - Capability discovery — Pattern for helping users discover what an AI can do: slash command menus, rotating hints, contextual tips. https://stepikin.com/llm-ux-patterns/capability-discovery/ - Citations & sources (RAG) — How to design inline citations for RAG: markers, source cards, quoted snippets, freshness and failed sources. https://stepikin.com/llm-ux-patterns/citations-and-sources/ - Showing uncertainty — How to show AI confidence honestly: claim-level uncertainty, reasons and verify actions, without fake percentages. https://stepikin.com/llm-ux-patterns/confidence-and-uncertainty/ - Context chips & attachments — How to show what an AI model can see: attachment chips, @-mentions, and a context-window meter. https://stepikin.com/llm-ux-patterns/context-attachments/ - Error & refusal states — How to design error and refusal states for LLM products: network drops, refusals with alternatives, over-limit input. https://stepikin.com/llm-ux-patterns/errors-and-refusals/ - Evals as a product surface — Design pattern for LLM evaluation dashboards: scores across versions, pass thresholds, regression detection and failing cases. https://stepikin.com/llm-ux-patterns/evals-dashboard/ - Feedback: thumbs + reasons — How to design feedback for AI answers: thumbs, reason chips, optional comment and consent to share. https://stepikin.com/llm-ux-patterns/feedback-thumbs/ - Hallucination guardrails — Pattern for visible hallucination guardrails: claim checks against sources, unsupported-claim flags and a strict grounded mode. https://stepikin.com/llm-ux-patterns/hallucination-guardrails/ - Human-in-the-loop confirmation — How AI agents should ask for permission before irreversible actions: what to show, which buttons, when not to ask. https://stepikin.com/llm-ux-patterns/human-in-the-loop/ - Plan preview — Design pattern for showing an AI agent's plan before execution: editable steps, risk labels, cost and time estimates. https://stepikin.com/llm-ux-patterns/plan-preview/ - Data & privacy disclosure — AI privacy UX: detect personal data before sending, offer redaction, and show where data goes and how long it is kept. https://stepikin.com/llm-ux-patterns/privacy-disclosure/ - Prompt suggestions — How to design prompt suggestions and follow-up chips for AI products: when to use them, anatomy, do & don't, a live demo and implementation notes. https://stepikin.com/llm-ux-patterns/prompt-suggestions/ - Prompt templates with variables — Design pattern for prompt templates with variables: turn a proven prompt into a form with inline fields. https://stepikin.com/llm-ux-patterns/prompt-templates/ - Rate limits & quotas — Design pattern for AI quotas and rate limits: usage meters, early warnings, reset timers and fallback models. https://stepikin.com/llm-ux-patterns/rate-limits-and-quotas/ - Stop & regenerate — How to design stop, continue and regenerate controls for LLM responses, with version history. https://stepikin.com/llm-ux-patterns/stop-and-regenerate/ - Streaming responses — How to design streaming AI responses: status steps, token streaming, first-token latency and a stop control. https://stepikin.com/llm-ux-patterns/streaming-responses/ - Structured output & cards — Design pattern for rendering AI answers as structured cards instead of prose: comparisons, entities, actions. https://stepikin.com/llm-ux-patterns/structured-output/ - Tool-call transparency — Design pattern for showing AI tool and function calls: collapsed summaries, arguments, results, and permission prompts. https://stepikin.com/llm-ux-patterns/tool-call-transparency/ - Undo for agent actions — Pattern for making AI agent actions reversible: activity feed, timed undo, undo the whole run. https://stepikin.com/llm-ux-patterns/undo-agent-actions/ - Voice input for AI — Voice input pattern for AI products: listening state, live interim transcript, edit-before-send. https://stepikin.com/llm-ux-patterns/voice-input/ ## Templates for founders (free, PDF and Markdown) Index: https://stepikin.com/templates/ - AI feature scoping: code, prompt, RAG or fine-tune?: https://stepikin.com/templates/ai-feature-scoping-worksheet/ - The AI MVP brief: one page that decides what you build first: https://stepikin.com/templates/ai-mvp-brief/ - AI product launch checklist: 42 checks before real users touch it: https://stepikin.com/templates/ai-product-launch-checklist/ - How to choose a design and dev partner for an AI product: https://stepikin.com/templates/ai-product-partner-scorecard/ - The 3-minute investor demo: a script for AI products: https://stepikin.com/templates/investor-demo-script/ ## Guides, tools and blog - AI product cost and timeline estimator: https://stepikin.com/estimate/ - Free llms.txt generator: https://stepikin.com/llms-txt-generator/ - FAQ — hiring a designer + developer for an AI product: https://stepikin.com/faq - What is a solo+AI studio: https://stepikin.com/solo-ai-studio/ - Best AI product design studios in 2026: https://stepikin.com/best-ai-design-studios/ - Design + code for AI startups: https://stepikin.com/ai-startups - Blog: https://stepikin.com/blog - Build diary — native desktop SaaS with AI (Tauri 2, 44 releases): https://stepikin.com/blog/engiboard-6-days/ - I tried to predict outbreaks with AI and failed twice — what I built instead: https://stepikin.com/blog/vigilo-honest-radar/ ## Contact - Book a 30-minute call: https://stepikin.com/book/ - Email: hello@stepikin.com - Telegram: @uxuialex (https://t.me/uxuialex) - LinkedIn: https://www.linkedin.com/in/alekseystepikin/ - Behance: https://www.behance.net/alexstepikin - Website: https://stepikin.com/