Overview of AI Skills with Matt Pocock
In this episode, Gergely Orosz talks with Matt Pocock about his unusual path from voice coach to developer-educator, and how that background shaped his approach to teaching software engineering and building AI workflows. The conversation centers on Matt’s popular AI “skills” — especially Grill Me and Wayfinder — and why classic software engineering fundamentals, rather than flashy new AI ideas, have become even more important in the age of agents.
Matt Pocock’s path into tech
From voice coaching to software
- Matt spent about six years as a voice coach / singing teacher before becoming a developer.
- He taught:
- singing and accents
- voice and public speaking
- Shakespeare / drama-school performance
- consulting teams how to speak and present better
- He became a developer largely because he wanted a remote-friendly career and didn’t want to stay in London.
Early developer career
- He self-taught JavaScript by building tools for his coaching work, including:
- flashcard apps
- a very ambitious voice-analysis tool
- He got his first dev job in 2017, then moved through several agencies and later into open source and developer education.
- A big turning point came through TypeScript and XState, where he became heavily involved in advanced type work and the XState core team.
- He later joined Vercel, where he began building educational content more seriously and eventually launched Total TypeScript.
Total TypeScript and the move into education
Why teaching worked
- Matt’s earlier career as a teacher gave him a major advantage:
- he was comfortable explaining complex ideas
- he could interview well
- he knew how to guide people through learning step by step
- He noticed a huge response to his short TypeScript tips on Twitter, which convinced him there was a market for deeper educational content.
- Total TypeScript took off quickly and became a major business.
His business model
- Matt prefers a straightforward model:
- sell products and courses
- keep a strong refund policy
- avoid sponsored content where possible
- He sees education as a way to create high-leverage work that helps people and supports his family life.
How AI changed everything
Knowledge is cheap; wisdom is scarce
Matt’s core view is that AI has made knowledge much easier to access, but not wisdom.
- AI can explain syntax and generate code.
- But it still cannot replace:
- judgment
- prioritization
- tradeoff thinking
- long-term architectural decisions
He draws a distinction between:
- tactical programming: immediate coding and implementation
- strategic programming: system design, workflow design, and decision-making over time
His argument is that AI has mostly eaten tactical work, which makes strategic thinking more valuable than ever.
Matt’s AI skills: Grill Me, Wayfinder, and more
Grill Me
A skill designed to make AI act more like a strong senior engineer.
What it does
- Interrogates you aggressively before coding.
- Forces you to answer questions about:
- scope
- requirements
- tradeoffs
- auth / permissions
- edge cases
- Helps the model understand your priorities and boundaries instead of improvising.
Why it works
- It reduces misalignment between human intent and agent behavior.
- It gets the model to challenge assumptions instead of just outputting code.
- It feels like being grilled by a thoughtful senior engineer.
Wayfinder
A more ambitious skill for larger, multi-session work.
Core idea
- Break big projects into:
- a spec / destination document
- many smaller tickets
- Use a “map” to track what is known, what is still foggy, and what needs more exploration.
What it’s for
- Planning larger pieces of work
- Research and prototyping
- Work that spans many agent sessions
- Keeping the agent oriented across long-running tasks
Why these skills spread
- They’re simple to understand.
- They encode practical workflows, not abstract theory.
- They reflect how Matt actually works, which makes them usable and easy to adapt.
Why old software books matter more in the AI era
Matt repeatedly returns to classic books and ideas because they still map well to agentic coding.
The books he recommends
- The Pragmatic Programmer
- A Philosophy of Software Design — John Ousterhout
- Domain-Driven Design — Eric Evans
Concepts he highlights
- Software entropy: codebases naturally drift toward messiness
- Tracer bullets: build thin working slices that give fast feedback
- Vertical slices: integrate across layers early instead of building everything in isolation
- Deep modules: strong abstractions with real value
- Ubiquitous language: shared domain vocabulary between humans and code
His view is that these ideas were always good, but AI makes them even more important because agents are more likely than humans to generate confusion and technical debt.
Leading words and better prompting
One of the most interesting ideas in the episode is Matt’s use of leading words.
What they are
- Terms like:
- tracer bullet
- vertical slice
- deep module
- ubiquitous language
- These act as cues that steer the agent toward a better behavior pattern.
Why they matter
- They help align the model with the mental model of the developer.
- They compress communication.
- They tap into concepts that are likely already present in the model’s training data.
This is a major theme in the episode: good language creates better output.
Working with agents well
Clean codebases matter more now
Matt argues that agents are much better at working in:
- well-structured codebases
- clear domain models
- codebases with meaningful terms and clear boundaries
Why?
- Agents do not have true long-term memory.
- Each session is effectively a fresh start.
- So the codebase itself becomes the “memory” and the environment.
Observability and automated review
For teams, he recommends:
- measuring agent success and failure rates
- understanding where agents produce bad outputs
- using automated review agents to enforce standards
- improving the environment rather than only blaming the model
TDD still helps, but differently
Matt is cautiously positive on TDD:
- It’s less about human memory now.
- It’s more about forcing feedback loops and making the agent prove the code works.
- He prefers prompts like: “Provide evidence that this would fail without the change.”
Tools and workflow changes
Local vs cloud
Matt says he’s increasingly moving away from local-only setups.
Why
- Collaboration is easier in shared places like:
- Slack
- Discord
- cloud boxes
- Cloud environments make it easier to:
- keep resources always available
- schedule tasks
- let agents work while you’re away
- He likes the idea of a “day shift” for humans and a “night shift” for agents.
Main takeaways
- AI makes tactical coding faster, but strategic thinking more important.
- Old software engineering books still contain some of the best guidance for working with agents.
- Prompting is partly a language problem: the words you use shape the model’s behavior.
- Good AI workflows are about structure:
- specs
- tickets
- feedback loops
- review
- observability
- Clean, domain-rich codebases help agents work better.
- The best developers in the AI era will likely be introspective, curious, and systematic.
Recommended reading from Matt
- The Pragmatic Programmer
- A Philosophy of Software Design
- Domain-Driven Design
- especially the sections on ubiquitous language and domain modeling
Closing thought
Matt’s bigger message is that AI doesn’t replace software engineering fundamentals — it exposes which ones actually matter. The better you are at defining problems, naming things, structuring work, and designing feedback loops, the better AI tools will perform for you.