12 Ways to Use Claude So Well It Feels Illegal (Tutorial)
Sabrina Romanov presents 12 productivity hacks for using Claude and other AI tools effectively, ranging from foundational prompt engineering to advanced techniques like MCP integration and skill stacking. The tutorial progresses from basic concepts like structured prompting to sophisticated workflows involving autonomous AI agents that can execute real-world tasks. The overarching message is that consistent hands-on practice separates true AI power users from passive consumers.
Summary
The video opens by framing two types of AI users: those afraid of being replaced and those actively building businesses with AI tools. Sabrina Romanov, who sold her AI company for millions, presents 12 ordered hacks that build upon each other from foundational to advanced.
Hack 1 covers prompt engineering using a simple four-part template: define the AI as a top 0.1% expert in a field, state the task, provide context, add constraints, and append 'ask me clarifying questions one at a time until 95% confident.' This replaces expensive courses and yields dramatically better outputs than vague prompts like 'make it better.'
Hack 2 reframes AI from a 'glazing partner' that validates ideas into a sparring partner that challenges them. Sabrina recommends prompting AI to identify blind spots, weak assumptions, missing data, and arguments against your own plans — treating it like a candid advisor rather than a yes-man.
Hack 3 encourages using AI as a 24/7 tutor instead of waiting for tutorials or asking creators for help. Dropping screenshots, describing confusion, and asking AI directly builds self-sufficiency and leverages AI's infinite patience and breadth of knowledge.
Hack 4 introduces Claude 'skills' — reusable, context-rich task definitions analogous to custom GPTs but more composable. Sabrina describes her YouTube script workflow broken into research, outline, writing, editing, and fact-check skills, and recommends having Claude generate skills automatically from conversations using a self-reflection prompt.
Hack 5 addresses memory, specifically the Claude.md file in Claude Code — a plain-text file Claude reads at session start containing brand voice, architecture decisions, goals, and constraints. Sabrina recommends continuously updating it by prompting Claude to extract key learnings after each conversation, rather than setting it once and forgetting it.
Hack 6 highlights repetition: a Google research study reportedly found that repeating key instructions during a session improves AI output quality. Sabrina runs a shortcut at the start of every session forcing Claude to re-read her Claude.md file, and spams a 'minimal changes' shortcut multiple times before any code implementation.
Hack 7 focuses on planning, specifically spending 80-90% of session time in plan mode before switching to execution. This involves using sparring and prompt engineering techniques to rigorously validate the approach, and then closely monitoring Claude's execution to abort if it deviates from the agreed plan.
Hack 8 explains MCP (Model Context Protocol), which connects AI to real-world tools like Stripe, Gmail, Notion, Airtable, and Slack — bridging the gap between AI as a consultant giving advice versus an employee actually executing tasks. Sabrina built her own MCP server for her content tool Potato, enabling TikTok scraping, visual generation, and social media scheduling from within the AI interface.
Hack 9 covers stacking skills with MCP, describing this combination as 'the closest approximation to an AI employee.' Skills define what to do and how to do it well; MCP provides the tools to actually execute. Her cross-post skill exemplifies this: it finds TikTok drafts in Google Drive, transcribes them, writes platform-specific captions, generates visuals, schedules via Potato MCP, and logs everything in Airtable — all in one command with flexible exceptions handled through natural language.
Hack 10 covers mobile apps, particularly Claude Code's Remote Control feature, which lets users monitor and approve Claude's actions on their local computer from their phone without enabling dangerous auto-permission modes.
Hack 11 introduces GitHub for Claude Code users as a way to safely experiment via branches without risking the main project, and to version-control Claude settings and skills.
Hack 12 emphasizes putting in reps — actually doing the work rather than passively consuming tutorials. Sabrina uses the gym analogy repeatedly: watching workouts produces no fitness results, just as watching AI tutorials without hands-on practice produces no skill development.
Key Insights
- Sabrina argues that a four-part prompt template — defining AI as a 0.1% expert, stating the task, providing context, adding constraints, and asking for clarifying questions one at a time until 95% confident — captures 99% of what expensive prompt engineering courses teach.
- Sabrina claims that combining Claude skills (which define what to do and encode preferences) with MCP connections (which provide access to real tools like Stripe, Airtable, and Google Drive) creates 'the closest approximation to an AI employee' she has ever experienced — one that can adapt to exceptions stated in natural language, unlike rigid workflow automation tools like n8n or Make.com.
- Sabrina describes a Google research study finding that repeating the same instructions to AI during a session actually results in better answers, which she says validates her practice of spamming a 'minimal changes' shortcut multiple times before any code implementation to keep Claude focused on her most important constraints.
- Sabrina contends that most AI users make the mistake of using AI as a 'glazing partner' that validates any idea, and that switching to a sparring-partner dynamic — asking AI to identify blind spots, weak assumptions, and missing data — fundamentally changes the quality of decisions made.
- Sabrina recommends continuously updating the Claude.md memory file by prompting Claude after each conversation to extract and log the most important learnings, rather than writing it once and leaving it static — arguing this creates highly tailored, evolving documentation that compounds in value over time.
Topics
Full transcript available for MurmurCast members
Sign Up to Access