Practical AI Coding Workflow; 3 Rules Before You Build; AI Coding Showdown
Issue #87 Bytes
đ± Dive into Learning-Rich Curation with groCTO —ïž
Article of the Week â
âBad codebases make bad agents. If you have a garbage codebase you're going to get garbage out of the agent working in that codebase.â âMatt Pocock
Full Walkthrough: Workflow for AI Coding â Matt Pocock
Weâre changing up this format for this week as we found an interesting workshop video that all you AI enthusiasts will find very practical!
Matt Pocock has spent the last six months stress-testing AI coding workflows: using battle-tested old school principles and tacking it onto AI / claude workflows. In his hands-on workshop from a live conference, he walks through his full development lifecycle end-to-end, from a vague Slack message to production-ready code shipped by an autonomous agent.
1. The smart zone is finite, design around it.
LLMs degrade in quality as context grows. Pocock puts the reliable ceiling at around 100k tokens regardless of advertised context window size, and every workflow decision he makes is optimised around staying inside it. That means short, focused sessions with clean context resets rather than endlessly compacted threads. The 1 million token context window doesnât change this as it merely ships you more âdumb zoneâ. Teams that ignore this are burning tokens and wondering why output quality is inconsistent. The fix isnât a better model. Itâs smaller, cleaner tasks.
2. Get alignment instead of planning with the âgrill meâ skill.
His most distinctive move is refusing to let the AI produce a plan until it has interrogated you thoroughly first. His /grill-me skill forces the agent to interview you relentlessly â sometimes 40 to 80 questions â before anything is written down.
The goal is to produce what he calls a shared design concept: getting on the same wavelength as the agent that will implement your work.
He draws on Frederick Brooks here (the author of the Mythical Man-Month and The Design of Design), the idea that everyone building something together needs to hold the same mental model of what theyâre making. Skipping this step and going straight to specs-to-code is, in his words, vibe coding by another name. Youâre ignoring the code and hoping the document saves you. It doesnât.
3. Vertical slices, not horizontal layers.
Left to its own devices, AI codes layer by layer, e.g. database first, then API, then frontend.
That means you get no integrated feedback until the very end of the plan. Pocock structures work instead as thin tracer bullets that cut across all layers from the start, so the agent can run a full feedback loop after every issue rather than after phase three.
This also unlocks parallelisation: once issues are structured as a directed acyclic graph with explicit blocking relationships, multiple agents can work simultaneously on independent branches and merge cleanly. A sequential multi-phase plan can only be picked up by one agent. A well-structured kanban board can be picked up by several.
4. Deep modules are the ceiling on agent quality.
Drawing on John Ousterhoutâs A Philosophy of Software Design, Pocock argues that the structure of your codebase is the single biggest lever on agent output quality. Shallow modules with lots of small, tightly-coupled files are hard for agents to navigate and nearly impossible to test well.
Deep modules with small interfaces and rich internals give agents a clear surface to work against and draw clean test boundaries around. If your AI is consistently producing mediocre output, the problem is probably that youâve handed an agent a codebase that experienced human engineers also struggle to work in cleanly.
The planning frameworks senior engineers already know translate almost directly into better agent output. The teams struggling with AI are skipping the discipline that made good software before the agents arrived.
Other highlights đ
The Critical Shift In What Differentiates Great Leaders
Being the smartest person in the room used to be a career strategy. Yue Zhaoâs brings a delightful perspective on how most senior leaders are dangerously over-invested in the one type of intelligence AI just made cheap.
Her framework splits human intelligence into three centers: Head, Heart, and Gut. Most high-performing leaders have spent their entire careers developing exactly one. AI just made that one the table stakes.
1. The Head is now a commodity. AI has read more case law, medical literature, and financial filings than any human ever will. Intellectual horsepower, ie. the thing that got most senior engineers into the room, is now available to everyone on an instant. Well, mostly.
2. The Heart is where humans still have the edge. AI can fake empathy. It cannot feel it. When a team is demoralized after a reorg, when two opinionated people are locked in conflict, when someone needs a hard truth delivered in a way they can actually hear are the moments that require a human who is genuinely in touch with whatâs happening emotionally and skilled enough to do something useful with it.
3. The Gut is the final frontier. Ethical judgment in genuinely novel situations. Where no published framework quite fits and someone has to stand for something. AI can reflect the ethics humans have written down. It cannot tell you what is right when the rules run out.
The skills that got you to staff or director are still necessary. Theyâre just no longer sufficient. But the skills that will define your next decade as a leader have nothing to do with how well you can out-think a machine and everything to do with what only a human can do.
3 constraints before I build anything
Products fail when they lack a forcing function to stay honest about what they were building. Jordan Lord shares his practical framework on how to triage his idea economy with a tight, practical rulebook that maps cleanly onto how engineering leaders should be evaluating what their teams take on.
1. One page or it doesnât get built. If you canât describe the idea in a single page youâre not ready to build it. The one-pager is the north star: non-negotiable, precise, and lean.
It also becomes the tie-breaker for every conflict that surfaces during development. If the thing youâre arguing about isnât in the one-pager, itâs either not worth fighting over or the one-pager needs to be updated. Critically, if you canât fill the page without padding it with fluff, thatâs a signal to stop and do more research; not to start building.
2. The core tech must be separable from the product. Every idea worth building should produce a reusable piece of technology that outlives the product itself. A method, a library, a tool, a methodology. The examples here are instructive: Linus Torvalds built Git to improve the Linux kernel workflow. HashiCorp built HCL. Google built Kubernetes. Products pivot constantly. Core tech compounds. If your idea doesnât generate separable leverage, it probably isnât high enough value to justify the investment.
3. One defining constraint must shape the product. Every product needs a constraint that is visible to the user at all times: something that gives the product its identity and limits scope by design. Minecraft is made of blocks. IKEA is flat-pack. The constraint is a filter that collapses the decision space and forces the team to solve the problems in novel ways. Without a deliberate constraint products tend to bloat out of control or worse: seek constraints from market factors after having been built.
Constraints arenât the enemy of good work. Theyâre often what makes it possible.
Choosing the Right AI Coding Tool
AI coding tools are beginning to shape how teams think, review, debug and even structure engineering work. But the real differences are less about features and more about workflows. Some optimise for speed inside existing habits, while others reshape development around deeper AI collaboration and reasoning.
Whatâs becoming clear is that teams are no longer looking for one universal assistant. Theyâre starting to match tools to different kinds of work, from rapid iteration to large-scale refactoring and architectural thinking. This piece breaks down those trade-offs in a practical, grounded way.
Find Yourself đ»
Thatâs it for Today!
Whether youâre innovating on new projects, staying ahead of tech trends, or taking a strategic pause to recharge, may your day be as impactful and inspiring as your leadership.
See you next week, Ciao đ
Credits đ
Curators - Diligently curated by our community members Denis & Varun
Featured Authors - Matt Pocock (c. AI Engineer, youtube), Yue Zhao, Jordan Lord
Sponsors - This newsletter is sponsored by Typo AI - Engineering Intelligence Platform for the AI Era.
1) Subscribe â If you arenât already, consider becoming a groCTO subscriber.
2) Share â Spread the word amongst fellow Engineering Leaders and CTOs! Your referral empowers & builds our groCTO community.
