Matt Pocock

A possible future: For the last 30 years, devs have made the sensible assumption that every time a human touches a codebase under time pressure, the codebase gets less maintainable. This is called software entropy. It's only possible to escape by continual monitoring and improvement. Or by employing ONLY 10x devs (impossible). The current generation of LLMs aren't currently good enough to escape software entropy. Without careful supervision, they make codebases worse. But soon, they will be good enough that they will make the codebase better each time they touch it. At that point, we won't need human review any more. Codebases will be better as a black box. We'll review inputs and outputs. Before Opus 4.5, models produced software entropy at an alarming rate. Opus 4.5 was the first time I felt like the entropy was manageable. Soon, there might not be any entropy at all.

History of AI Coding Vibe Coding: don't look at the code. Go fast. Only for prototypes Plan Mode: increase code quality by planning first Multi-Phase Plans: use a single plan across multiple context windows Kanban: Keep a kanban board of tasks, run the agent in a loop

Ever wondered why the longer your chat grows, the stupider the AI becomes? It might be because you don't understand how context windows work. Let me enlighten you:

I just nearly fell into a 50-year-old trap: waterfall. "I'll plan everything upfront, feed the spec to the LLM, ship it all." But specs→code isn't like a compiler. Some questions only get answered in code. Note to self: Validate the risky parts first.

How it started: Swarms, Orchestrators, Agent meshes, Multi-agent frameworks, Hierarchical agent systems, Agent-to-agent protocols How it's going:

Two ways to make your 'Plan Mode' plans better: - Put "sacrifice grammar for the sake of concision" in your rules - Put "list any unresolved questions at the end, if any" at the end Means the first thing you see is a hyper-concise list of immediately actionable questions.

Thinking about a free Ralph tutorial Start with the basic loop, then layer on top: - Containerization - Feedback loops - Testing - Formatting - Linting/types - Skills (for steering) - To plan or not to plan? All using it to build an actual production app. WDYT?

I've been using the crap out of Claude Code recently And lots of you have been asking me my tips So, here goes - 10 minutes of me using Claude Code to implement a feature.