In his talk at AISN 2026, Andrej Karpathy, the former Head of AI at Tesla, revealed how most people are prompting AI models like Claude incorrectly. To build ten times faster, Karpathy utilizes a structured three-layer framework: the Spec, the Verifier, and the Environment. This systematic approach shifts users away from lazy prompting toward modern AI engineering, leveraging human context to direct raw computational power.

Layer 1: The Spec

AI models excel at measurable computation but fail at context-driven, common-sense scenarios (such as walking to a car wash 50 meters away). To bridge this gap, Karpathy advises against high-level planning modes in favor of detailed, co-designed specs:

• Uncover the Goal: Do not just prompt for a task. Instruct Claude to interview you to identify the core goal or the specific decision your project is meant to drive.
• Agile Specking: Avoid “waterfall” execution where you ask the AI to complete a massive project all at once. Instead, break the project into small, compartmentalized phases with clear checkpoints.
• Precision and Critical Thinking: Be highly precise to prevent the AI from making assumptions. Force the model to explicitly ask you to verify key decisions before moving forward.

Layer 2: The Verifier

Unlike humans (whom Karpathy compares to motivated “animals”), AI acts like a “robot librarian” or a statistical simulation. It has no intrinsic drive and will confidently make things up when it lacks information. To maintain quality control, you must optimize your verification processes:

• Precise Evaluation Criteria: Define exactly what a successful output looks like upfront (e.g., specific formatting, sections, or logic) before the AI starts building.
• Multi-Model Criticism: Use a second AI model or specialized tool (such as CodeEx in Claude Code) to act as a critic and review the output of the first model.
• Integrate External Signals: Connect your AI workspace with live systems, deployment environments, or historical data to verify outcomes with absolute certainty.

Layer 3: The Environment

Instead of starting from a blank chat template every time, you must build a persistent, optimized workshop for your AI agents to live in:

• Custom claude.md Files: Establish a system file containing repository structures, routing rules, and mandatory verification instructions that the AI automatically digests.
• LLM Knowledge Base: Organize a structured folder system of proprietary data on your local machine to serve as a specialized training library.
• Custom Skillsets: Create and iterate on step-by-step custom handbooks for repeated tasks.
• Hard Guardrails: Define clear boundaries at the system level (Always Do, Ask First, Never Do) to restrict tool-use and ensure the AI never modifies critical files without permission.

Conclusion: Outsource Thinking, Keep Understanding

Karpathy’s ultimate takeaway for navigating the future of AI is clear: “You can outsource your thinking, but you can’t outsource your understanding.” The three-layer framework works because it relies entirely on your strategic direction, goal setting, and evaluation. Mastery of AI comes down to how effectively you can direct the machine using your human context and understanding.

Mentoring question

How can you transition your current AI workflows from a ‘waterfall’ approach to an ‘agile’ system that leverages co-designed specs and multi-model verification?

Source: https://youtube.com/watch?v=7zZy1QTvokM&is=Y8bMmJch6tlYX2pR