Understanding Coding Productivity With AI Today

Practical podcast interview with Matt Pocock (AI engineering educator and creator of popular agent skills).

The core message is that the model is important, but the harness (your skills/prompts/procedures + environment + workflow) matters just as much or more.

XAI bought Cursor because it is a top level harness and they have the data and proven product market fit for AI coding and AI harness.

Screenshot

Tactical vs Strategic Programming

AI has mostly eaten tactical coding (writing boilerplate, fixing small bugs, implementing features from specs). The real differentiator now is strategic work — architecture, long-term planning, product thinking, and maintaining high-quality codebases.

The Harness is greater than Raw Model Power.

Matt repeatedly stresses building robust systems around the AI rather than just using the latest model in a basic chat. He shares his open .claude skills repo as an example.

Skills Design and Procedures (user-invoked, more controllable) vs Abilities (model-invoked).

He prefers procedures for reliability.

His famous Teach skill creates personalized, adaptive learning plans (HTML lessons, progress tracking, etc.).

Agentic Setup Recommendations

Uses Claude Code (with Opus 4.8 at medium effort) + Sandcastle for safe/sandboxed AFK agents.

Prefers queues (GitHub issues + parallel GitHub Actions) over real-time infinite agent loops.

Strong emphasis on human-in-the-loop and AFK (away-from-keyboard) work to scale output.

Matt acknowledges Fable’s impressive capabilities (autonomous browser use to fix API issues) but is cautious — he’s still mostly on Opus 4.8 and hasn’t fully switched. He highlights risks like security concerns and models disappearing or becoming expensive.

Future of Developers is more focus on product design, orchestration, delegation, and fundamentals. Good codebases and processes become even more valuable.

Why CURSOR Matters for xAI and Can Become Relevant and Competitive

Strengths are Vertical integration win

IDE → Models → Compute → Git platform. Hard for pure model labs to compete.

Data moat of Millions of real developer sessions.
Product-market fit is proven. Cursor already has strong developer love and usage.
Internal tooling SpaceX/xAI get world-class AI agents for rocket/software engineering at massive scale.
Compute advantage realized Colossus finally gets a high-volume, high-signal application.