How to Code Smarter with AI (and Avoid the Biggest Mistakes) | Elie Schoppik (Head of Education)

Listen now on YouTube, Apple Podcasts, or Spotify.

👤 Guest Bio

Elie Schoppik leads Technical Education at Anthropic, building Claude Code’s learning-first workflows and Anthropic Academy to teach people how to verify AI outputs (not just consume them).

Before Anthropic, he founded early ed-tech products and Rhythm School, a Bay Area coding bootcamp that trained thousands of developers. Elie’s work sits at the intersection of engineering and pedagogy; designing exercises, evals, and prototype workflows that teach verification, build durable judgment, and make AI safe and useful in the real world.

PRESENTED BY AUTOSKILLS

Take your team from AI-curious to AI-ready in months days

Autoskills helps teams to go from AI-curious to AI-ready with:

→ AI acceleration sprints
→ Fractional AI automation engineers to build AI workflows
→ Custom AI transformations

Teams that work with Autoskills cut hours of repetitive work, identify high-ROI use cases, and leave with the confidence (and playbook) to scale AI responsibly.

Limited to 3 clients per quarter - book a free AI Readiness Audit today!

🎙Episode Intro

When ChatGPT landed, “fast answers” replaced judgment for many learners, creating a gap between getting an answer and knowing if it’s correct. Elie’s work is about closing that gap: building workflows where AI is the teacher’s assistant, not the oracle. He gives us practical rules, concrete design patterns (examples-first prompts, end-state screenshots, evals), and ready-to-run exercises every team can use to build AI fluency.

⏱ What’s Covered

• (02:29) Elie’s background and leading technical education at Anthropic

• (03:41) How AI education differs from coding

• (07:15) Anthropic’s products: Cloud Code, MCP, and model actions

• (17:36) Why examples and context matter in prompting

• (18:45) Context engineering: narrowing tasks to get better model output

• (19:59) Core topics to study beyond tool use

• (32:26) Role of education in helping companies join the “successful 5%”

• (35:37) Hackathons as a catalyst for AI adoption in organizations

• (44:59) Dopamine hits, quick wins, and why they matter in AI learning

• (49:03) Can mastering AI tools reduce “AI anxiety”?

• (50:47) Balancing speed vs. responsibility in an AGI timeline

• (52:29) Anthropic Academy resources for technical + non-technical teams

• (55:17) What excites Elie most about the future of AI learning

• (56:03) How parents can prepare kids for an AI-driven future

💡 Key Takeaways

• Teach verification, not shortcuts. Build learning modes that require users to answer follow-ups, explain reasoning, or run checks. Don’t let the model hand off the whole answer.

• Start with the building blocks. Teach high-level context engineering (MCP, data connectors, subagents) so teams understand tradeoffs even if they never run a server.

• Prototype with an end state. Give the model concrete examples or a screenshot of “what good looks like” and iterate. It’s often faster and more reliable than vague instructions.

• Ship evals early. Define unit tests/evals for any LLM workflow to avoid brittle deployments and false confidence.

• Short sprints beat courses. Run concentrated, hands-on learning + immediate application to capture momentum and build confidence.

• Make the loop self-improving. Use Cloud Code + Claude to generate assessments (quizzes), evaluate responses, and retrain the learning flow — that closed loop is a force multiplier for fluency.

Exercises you can run this week

• 90-minute “Play Sprint” — pick a painful recurring task, give Claude a single “end state” example, and iterate until you get a working prototype. (Use Cloud Code for integrations.)

• Eval the evals (45 min) — for one LLM task (summarize customer emails), write 5 unit tests that define success/failure and run them against model outputs. If none exist, you’re building on a house of cards.

• Make-a-quiz loop (60 min) — ask Claude to create a short quiz on an internal topic, have Cloud Code assemble context (Slack, Drive snippets), run the quiz against a colleague, and let Claude grade + explain mistakes.

📚 References & resources

• Anthropic — Research lab building safer, more helpful LLMs (helps you rely on higher-quality assistant outputs).

• Anthropic Academy (courses & labs) — Free hands-on courses and labs to build AI fluency (helps teams get practical, job-ready skills fast).

• Model Context Protocol (MCP) — Open standard for wiring models to data/tools (Gmail, Drive, Slack). Powers reliable model ↔ data integrations. (Helps you connect LLMs to real systems safely).

• Claude Code / Claude desktop — Agentic coding + Cloud Code tooling for prototyping and executing tasks (helps teams prototype automations and integrate with existing stacks).

• Constitutional AI (research paper, 2023) — Training method that uses a “constitution” to make outputs helpful, honest, and harmless (helps you align model behavior to safety goals).

🔗 Where to Find Elie

• Twitter/X: @eschoppik

• LinkedIn: Elie Schoppik

👉 If you found this episode useful, share it with a teammate or founder navigating AI adoption.

And subscribe to AI Ready for weekly lessons on how leaders are making AI real at scale.

Until next time,
Haroon