Tech • AI • Robotics • Game

ENFR

How to Actually Code with AI (ft. Matt Pocock)

5/10
AINetworkChuckNovember 20, 2026 at 05:25 PM2:02:35
Audio player

⏱ ≈ 3 min read · 2 h 3 min video — 2 h saved

TL;DR

AI is making coding faster and more accessible, but effective software development still depends on humans providing strategy, structure and shared language.

KEY POINTS

Tactical work is being automated

Matt Pocock, known for earlier work in TypeScript and now for AI coding workflows, argues that modern models are highly effective at short-term implementation. They can generate code continuously and handle much of the tactical work that once filled a developer’s day. In his view, refusing to use AI for that layer of work is increasingly hard to justify because it is cheap, fast and often good enough.

Strategy remains the human advantage

The main weakness in current coding agents is long-term thinking. AI systems optimize for the task in front of them, not for maintainability months later, which can leave projects brittle as they grow. That is why experienced engineers still tend to produce stronger products: they know what to ask for, how systems should evolve and which trade-offs matter over time.

Pure vibe coding breaks down

Pocock draws a sharp line between AI-assisted development and so-called vibe coding in its strictest sense, where code is treated as an unimportant byproduct and barely reviewed. That approach, he says, causes quality to decay as each new change is optimized only for the present task. Since code is the environment the agent works inside, a neglected codebase gradually becomes harder for both humans and models to navigate.

Shared language is a core discipline

A central recommendation is to build a clear domain vocabulary for a project. Borrowing from domain-driven design, Pocock maintains a glossary of terms and a set of architectural decision records so the AI uses the same language as the database, front end and product logic. That reduces confusion, avoids multiple names for the same concept and makes large repositories easier to understand.

Agents should be forced to ask questions

Instead of giving a vague prompt and accepting the first output, Pocock uses a workflow that pushes the model to inspect the codebase and question the request before writing code. His tools include a Grill with Docs process that explores files, reads the project glossary and interrogates the user about requirements. The goal is to close the communication gap that often leads agents to add features that were never intended.

Context size is a hidden quality risk

Pocock argues that models become less reliable as more material is packed into context. Even with very large token windows, he prefers to keep agents in a smaller “smart zone,” saying judgment often starts to degrade well before the theoretical limit. His practical threshold is around 150,000 tokens, after which latency rises, costs increase and the model is more likely to make poor decisions.

Big features should be broken into specs and tickets

When a task is too large for one clean session, he recommends turning the discussion into a written specification, then splitting that spec into individual tickets. Each ticket becomes a focused implementation session followed by review against the original plan. This preserves strategic intent while still letting AI do the tactical execution.

Software engineering is becoming more accessible

Pocock contends that coding has always looked harder from the outside than it really is. With AI handling more implementation detail, newcomers can build useful products without traditional software backgrounds, provided they learn to care about quality, naming and system design. In that sense, the barrier is shifting from raw syntax knowledge to judgment and discipline.

Some work still benefits from stepping away from AI

For planning and curriculum design, Pocock says he often leaves the machine behind. He gathers audience signals from sources such as email, chat and social posts into a private wiki, but then prints learning goals on cards and rearranges them physically to shape lessons. That slower, human-paced process helps retain context in a way rapid model output often does not.

Knowledge work may be harder to automate than coding

Pocock sees a sharp distinction between coding and broader creative or planning tasks. Code can be tested and verified, making it ideal for AI assistance, while knowledge work often lacks clear feedback loops. That makes tasks like scripting, course planning or communication much harder to evaluate, even if AI can still support parts of the workflow.

CONCLUSION

The emerging model of software development is not code without engineers, but engineers and newcomers using AI to handle execution while humans guard architecture, language and intent. The immediate opportunity lies less in mastering one programming language than in learning how to direct agents without surrendering strategy.

Ask a question

More from AI