Practical AI use cases & workflows

Building persistent AI partnerships

I don't use AI as a tool — I've built a system where Claude is a thinking partner with persistent memory, calibrated behavior, and 14 integrated services. 61 named sessions. 150 memory files. 14 automated workflows. This is what happens when you take human-AI collaboration seriously.

61
Named sessions
14
Integrations
150
Memory files
30+
Client prototypes
what i build

Capabilities

Real products, real clients, real pipelines. Everything built through AI partnership — not prompting, but persistent collaboration with feedback loops.

Rapid Prototyping

Client brief to deployed demo in 2–3 days (previously 2–3 weeks). 30+ prototypes shipped for enterprise clients in automotive, construction, finance, and energy. Each auto-deployed via GitHub + Vercel.

VercelReactNext.jsTailwind

AI Education

Lecturer at NEWTON University (Prague, Brno, Bratislava). Teaching business students how to work with AI as a partner, not a search engine. Curriculum: prompt engineering, AI-assisted decisions, ethical deployment.

60 studentsscaling to 1203 campuses

E-Commerce Pipeline

Fully automated: raw supplement labels → AI-generated copy (13-block brief + 7 FAQ) → HTML product pages → Shoptet metadata → Figma cards → Canva banners. Label to live listing in hours.

7 products shippedShoptetFigmaCanva

Agentic Workflows

Claude-powered team bot with 6-layer security (allowlist, rate limiting, dedup, injection guard). Public tender scraper with auto-filtering. HR screening pipeline. Multi-entity invoicing. CRM integration.

14 workflowsMCPClickUpFakturoid

LLM Research

Empirical study: 200 outputs, 4 models (ChatGPT 5.2, Gemini 3 Flash, Claude Sonnet 4.5, DeepSeek V3), 5 coaching personas. System prompts boost quality by 21%. Discovered "over-prompting" — where personas degrade strong models.

p<0.001Cohen's d=0.87200 outputs

Persistent Memory System

150 memory files across 5 categories (identity, rules, projects, knowledge, feedback). Session lineage with handoff notes. Automated capture via shell hooks. 19 calibrated feedback rules from real collaboration.

150 files61 sessions19 rulesauto-capture
how it works

Workflow

Every session starts with context loading, continues with real work, and ends with automated knowledge capture. Two real examples:

session lifecycle
# 1. Context loads automatically
claude > Reading memory/LINEAGE.md
61 sessions loaded, last: #060 Locksmith
claude > Reading memory/MEMORY.md
150 memory files indexed
19 feedback rules active
14 integration configs loaded
# 2. Real work happens
human > build landing page, futuristic design
claude > Reading identity, projects, CV...
claude > ✓ Created, committed, deployed to Vercel
# 3. Auto-capture on session end
hook > activity.log ← session title extracted
hook > transcript → .staging/ for review
hook > Session #042 Builder captured
connected services

Tools & Integrations

14 services orchestrated through a single AI partner. Click any card to see what it does in practice.

ClickUp
MCP Server
Full project management through natural language. Tasks, sprints, time tracking, and client workflows.
  • Create tasks, assign priorities, set due dates
  • Track time on client projects automatically
  • Sprint planning and demo workflow management
  • HR pipeline: candidate tracking from screening to hire
Gmail
MCP Server
Read inbox, search messages, create context-aware drafts based on project state and conversation history.
  • Search and read emails by sender, subject, or date
  • Draft replies with full project context
  • Client outreach and follow-up automation
Google Calendar
MCP Server
Create events, check availability, coordinate schedules. Auto-notify team channels.
  • Create meetings with agenda and attendees
  • Find free time slots across calendars
  • Auto-notify team channels about new events
Figma
MCP Server
Design system management, product card generation, multi-screen layouts through code-driven Figma operations.
  • Clone and customize product card templates
  • Generate multi-screen interactive designs
  • Export assets for marketing pipelines
  • Design-to-code workflow with screenshots
Canva
MCP Server
Banner creation, asset management, brand kit access. AI-generated marketing materials from product data.
  • AI-generated marketing banners from briefs
  • Batch upload and organize assets
  • Access brand kits for consistent design
  • Export to PDF, PNG, JPG, PPTX
Gamma
MCP Server
AI-powered presentations, documents, and webpages. 105 themes including custom brand theme. AI image generation built in.
  • Sales decks and client presentations
  • Documents and webpages from briefs
  • AI images: Flux, Imagen, Ideogram, Leonardo
  • Export to PPTX and PDF
Fakturoid
REST API
Invoice generation for 3 business entities. Create, send, track invoices through natural language.
  • Multi-entity invoicing (3 business subjects)
  • Auto-generate from ClickUp task data
  • Send invoices directly via email
Mattermost
Bot API
Claude-powered team bot. Responds to @mentions, processes DMs, runs commands, bridges communication.
  • Natural language commands in team channels
  • 6-layer security (allowlist, rate limit, dedup, injection guard)
  • Persistent context from activity log
  • Cross-channel notifications and updates
GitHub
CLI
Every project auto-initialized with git, pushed to private repos, connected to Vercel for auto-deploy.
  • Auto-create repos from project scaffolding
  • Commit, push, and branch management
  • PR creation and issue management
Vercel
CLI
Zero-config deployment. Every commit triggers production build. Preview URLs for client review.
  • Auto-deploy from git push
  • Preview URLs for every branch
  • Connected to GitHub repos
Playwright
MCP Server
Browser automation: QA testing, web scraping, visual verification. Claude controls a real browser.
  • End-to-end QA testing of deployed demos
  • Web scraping for market research
  • Screenshot comparison and visual testing
  • HTML to PDF export pipeline
Context7
MCP Server
Live API docs for any library. Prevents hallucinated APIs by pulling current documentation into context.
  • Current docs for React, Next.js, Tailwind, etc.
  • Eliminates outdated API references
  • Version-specific documentation lookup
Shoptet
REST API
E-commerce product management. Listings, metadata, SEO-optimized descriptions, batch updates.
  • Product page creation with AI-generated copy
  • Metadata and SEO optimization
  • Batch product updates from templates
Mattermost Webhook
Webhook
Direct channel messaging for automated notifications. Calendar events, task updates, pipeline alerts.
  • Post formatted messages to any channel
  • Calendar event notifications
  • Pipeline completion alerts
system design

Architecture

A living system that learns. Four layers maintain persistent context across every session.

Lineage
61 named sessions with handoff notes. Each session inherits from the last.
Memory
150 files: identity, rules, projects, knowledge, feedback. Organized, not dumped.
Hooks
Auto-capture on session end. Activity log, transcript excerpts, context reinject on compact.
Calibration
19 feedback rules from real work. Structured disagreement, scope pushback, no recitation.
Session Start Load Lineage Load Memory Real Work Auto-Capture Next Session
research

Academic Work

Empirical, not theoretical. Data from real experiments, not blog posts.

The Effect of System Prompts on LLM Output Quality in Coaching Contexts

MSc. Thesis — NEWTON University, 2026
  • Controlled experiment: 200 outputs from 4 LLMs (ChatGPT 5.2, Gemini 3 Flash, Claude Sonnet 4.5, DeepSeek V3)
  • 5 coaching personas × 2 conditions (prompted vs. unprompted)
  • Result: system prompts increase quality by 21% (t(99)=8.65, p<0.001, Cohen's d=0.87)
  • Discovered "over-prompting" — explicit personas can degrade naturally strong models (Claude scored higher without prompts in 5 of 12 cases)

Teaching — "AI in Business"

NEWTON University, Prague · Nov 2025 – present
  • Teaching 60 business students (scaling to 120) how to use AI as a professional partner
  • Curriculum: prompt engineering, AI-assisted decisions, ethical deployment, hands-on tool use
  • Pilot AI subject at university — building the course from scratch
connect

Get in Touch

Martin Andrt

AI practitioner, educator, and product builder based in Prague. Lecturer at NEWTON University. Operations Manager at a digital agency. I build the bridge between how AI works and how people actually use it — then I teach others to cross it.

ai.andrt.martin@gmail.com LinkedIn