Svetlana Dergacheva

Class: AI educator · Agent builder · Product manager

I teach people to work with AI, and I build the guardrails that make AI agents safe to work with.

AI education Self-development Coding Harness engineering Security
Player bio

I am a product manager in EdTech who builds with AI every day. I write Claude Code skills, hooks and guardrails that keep agents honest: every claim has a source, nothing goes out without a human review, and nothing gets deleted by accident. I also teach. I wrote a course that takes analysts from zero to shipping AI-assisted code, AI lessons for kids aged 8–10, and Viola, a pixel game for learning Bulgarian.

Skill worlds

Five worlds I keep levelling up. Each one comes with proof you can open.

AI education

  • Author of a hands-on course for analysts: prompts, APIs, data apps, all built with an AI pair.
  • AI lessons for children aged 8–10: short, playful, no jargon.
  • AI-adoption workshops for product managers: from first prompt to daily habit.
  • Conference speaker on AI in analysis work.

Self-development

  • A second brain in plain markdown: decisions, sources and topics an agent can search.
  • Learning Bulgarian by building the game I learn with.
  • A kindness diary bot that helps children notice good deeds.
  • Every working day ends with a written result, not a feeling of being busy.

Coding

  • Viola: a browser game with 25 levels, worlds, races and a live player profile.
  • JavaScript, Python, HTML/CSS, SQL; Supabase, Cloudflare Pages, GitHub.
  • Spec-driven development: the spec comes first, the agent writes to it.
  • Telegram bots, data dashboards, generators and test bots that play my own game.

Harness engineering

  • Claude Code skills, slash commands, subagents and multi-agent workflows.
  • Hooks that check a file before it is written and save every session.
  • MCP servers wired to trackers, boards and browsers.
  • Mods for Claude Code itself, including my own dog that walks above the prompt.

Security

  • Guardrails for destructive commands: scan, show the list, then act; move to trash, never delete.
  • I audit third-party agent plugins line by line before I install them.
  • Secrets live in one locked file and never in chat, cards or commits.
  • Least privilege for agents: read by default, write only where I said so.

Quests completed

Things I built and shipped on my own time.

Game

Viola

A pixel adventure for adults learning Bulgarian. Vocabulary worlds, races, a castle, a player cabinet. English and Russian.

Play it →
Open source

PM second brain

A starter kit for product managers: a markdown vault plus six agent skills for inbox, calls, topics, specs and recall.

Open on GitHub →
Course

AI for analysts

Lessons that take an analyst from the first prompt to an API call and a working data app.

Link coming soon
Bot

Kindness diary

A Telegram bot where children write down the kind things they did today.

Private for now
Skills

Agent skills

Pre-flight checks for product work: claim check, objection map, UX audit with WCAG 2.2.

Coming soon on Agensi
Mod

Dog mod

A small pixel dog that walks above the Claude Code prompt. My first mod, built for fun.

Ask me for it

House rules

The rules my agents follow. They are also the rules I teach.

Raw sources stay untouchedNotes and summaries are built from them, never the other way round.
Every claim links to a sourceNo source means it is an assumption, and it says so.
Model output is a draftA person reads it before it counts.
Show contradictionsTwo sources disagree? Both stay on the page.
Nothing goes out by itselfThe agent prepares; a human sends.
Check both directionsIs what is printed true, and is anything important missing?
Inventory
Claude CodeSkillsHooks MCPSubagentsPython JavaScriptSQLSupabase CloudflareGitHubFigma ObsidianTelegram botsWCAG 2.2
Languages
  • RussianNative
  • EnglishWorking
  • BulgarianLevelling up

Press start to talk

Workshops, talks, AI coaching for teams, or a skill your agents are missing.