A Practical Roadmap for AI Proficiency in 2026

Jeff Su2026-06-09go watch the original →

the gist

Stop obsessing over prompt engineering and focus on providing high-quality context through Projects and interconnected AI systems to achieve better, more personalized outputs.

Master One Model and Optimize Defaults

Instead of jumping between models, choose one (ChatGPT, Claude, or Gemini) and use it exclusively to build transferable skills. Prioritize paid tiers, as the performance gap between free and paid versions is significant. Always manually select the most capable model available in the interface, as default settings often favor cheaper, less intelligent models. Use built-in memory features, such as Gemini's "Import Memory," to maintain continuity as you transition between tools.

Shift from Prompting to Contextualization

Effective AI usage relies on the "Outcome + Context" (OC) framework rather than complex prompt engineering. The model will infer roles, tone, and formatting if provided with sufficient context.

Use established frameworks: Reference specific methodologies like the "Pyramid Principle" to provide instant structural context.
Provide examples: Paste 2-3 previously approved documents or outputs to serve as a template for style, length, and tone.
Utilize Projects/Gems: Store recurring instructions, knowledge files, and project-specific memory in "Projects" (Claude/ChatGPT) or "Gems" (Gemini) to avoid repetitive setup.
Prefer Markdown: Use .md files for knowledge bases instead of PDFs, as they are easier for models to parse and cheaper to process.

Build Compounding AI Systems

Individual projects are often siloed. An AI system connects these silos to surface insights across different workstreams and allows for compounding feedback.

Cross-referencing: By centralizing data (e.g., health reports, workout plans, and financial data) into a single system like Claude "Cowork," the AI can identify patterns that individual projects miss, such as flagging a lack of cardio in a workout plan based on health checkup data.
Reconcile feedback: When editing AI-generated content, feed the final version back to the model with instructions to "reconcile" the changes. This forces the AI to analyze your edits and internalize your preferences for future tasks, reducing the need for manual instruction over time.