5 Claude Code skills I use every single day
TL;DR
Matt treats AI agents like decent engineers with no memory — his whole workflow is about encoding process into reusable Claude Code skills so agents can re-enter a repo and still behave consistently.
"Grill Me" proves tiny prompts can do huge work — the skill is only three sentences, but it routinely drives 30-50-question planning sessions, including one 16-question interview for a document-editing feature in his course video editor.
He turns vague ideas into durable artifacts before writing code — after alignment, "Write a PRD" explores the repo, interviews the user, and creates a GitHub issue with a problem statement, proposed solution, and agile-style user stories while keeping implementation details intentionally general.
"PRD to Issues" is about vertical slices, not layer-by-layer tasks — Matt breaks a PRD into a small kanban board of independently grabbable GitHub issues, using tracer-bullet-style slices to surface unknown unknowns fast; one example became just four issues with explicit blockers.
TDD is his most reliable quality lever for agents — the skill enforces a red-green-refactor loop, and he says that despite refactoring being the weakest step for LLMs, this pattern has most consistently improved Claude Code's output.
Architecture quality directly determines AI output quality — his weekly "Improve Codebase Architecture" skill hunts for shallow, fragmented modules, spins up 3+ sub-agents to propose competing refactors, and turns the winner into a GitHub refactor RFC because, as he puts it, garbage codebases produce garbage AI output.
The Breakdown
The core framing: AI as forgetful engineers
Matt opens with the idea that AI agents are basically a fleet of middling-to-good engineers who forget everything between sessions. That pushes the human role away from micromanaging code and toward building strict, repeatable processes — his Claude Code skills — that keep those agents from drifting.
"Grill Me": a three-sentence skill that forces real design thinking
His first daily skill is "Grill Me," and it's shockingly short: just three sentences telling Claude to relentlessly interview him until they reach shared understanding. He ties it to Frederick P. Brooks's The Design of Design and the notion of walking every branch of the "design tree" before committing to code; in practice, that meant Claude asked 16 straight questions about a document-editing feature, and sometimes these sessions stretch to 30-45 minutes with 30-50 questions.
From conversation to artifact with "Write a PRD"
Once the plan is clear, he switches to "Write a PRD," which formalizes the idea into a durable spec. The skill asks for the feature description, checks the repo so it doesn't hallucinate, interviews the user again, outlines major modules, and writes the final PRD as a GitHub issue — in his example, that included a problem statement, a split-pane UI proposal, and detailed user stories.
Turning the PRD into a kanban board of vertical slices
Next comes "PRD to Issues," which breaks the PRD into independently executable GitHub issues. Matt emphasizes vertical slices over horizontal ones, using the tracer bullet analogy: each issue should cut through the stack and flush out unknown unknowns early, especially around new integrations; one fairly involved PRD became just four issues with clear dependency links, and Claude later completed one by building a pure-function editing engine with 28 tests.
Why TDD is still his most dependable agent discipline
His fourth skill, "TDD," is much longer than the others because it forces a strict red-green-refactor loop. Before coding, the agent must confirm interface changes and target behaviors, and Matt says this matters because AI works better in codebases organized around a few deep modules with thin, obvious interfaces rather than lots of tiny scattered files; refactoring is still where LLMs stumble most, but the test-first loop has been his steadiest win.
A weekly architecture cleanup to keep future AI work sane
The final skill, "Improve Codebase Architecture," is less about building features and more about preventing entropy. It asks Claude to roam the repo and point out places where understanding requires ping-ponging across too many files, where pure functions were extracted just for testability while bugs actually live in orchestration, or where tightly coupled modules create risk — then it spins up three or more sub-agents to propose radically different interface designs and turns the winner into a refactor RFC.
The bigger lesson: write processes like a little book for humans
Matt closes by saying all five skills read like a mini markdown book of operating procedures. His broader point is simple: the best way to get quality from AI coding agents is to treat them like humans — except humans with the bizarre limitation that they wake up with no memory every time.