How I work
GymLogic is an open-source, MCP-native workout app. This page is the paper trail for how it gets built.
"The devs who succeed aren't the ones who delegate everything or nothing. They're the ones who fall back on engineering fundamentals."
Agent-driven ≠ vibe-coded. The fundamentals still apply — ubiquitous language, vertical slicing, TDD, real code review — and when they do, agents ship fast without trading away quality.
I'm not selling expertise. I experiment every day. What follows is a snapshot of one workflow as it stands today.
Self-critique
The grill
Before any code, before even a brief, every feature gets grilled. A
Cursor agent runs a grill-me skill — directly inspired by
Matt Pocock's grill-me — that asks one question at a time, recommends an answer with
rationale, and walks the design tree until every branch is resolved.
No multi-question dumps. No "what do you think?" with no opinion
attached. By the time it's done, half the bad ideas are already cut
and the survivors carry justifications.
Multi-agent critique
The critic in the room
I'm at the gym between sets. I notice a bug. I drop a 30-second voice memo on Telegram. Iris — a personal AI agent built on OpenClaw — listens, reads the repo, and files a fully-documented GitHub issue. Trivial fix? She assigns it to a Copilot agent. Bigger idea? I pick it up in Cursor next session and grill-me it from scratch.
Iris also reviews every brief I write in Cursor and pushes back on the gaps. Two agents critiquing each other's output catches things one author — agent or human — never sees alone. Same loop you'd ship for a team of humans, applied to a team of agents.
The workflow
Once the grill produces a brief, every feature runs through the same five steps. Each step has a Cursor skill that automates the process, and an artefact in the repo that proves it shipped.
-
Brief the epic
Walk the design tree with grill-me, then write up scope, out-of-scope, user stories, and acceptance criteria. The brief is what survives the grill.
-
Plan the tech
Translate the brief into architecture: key decisions, data model, component hierarchy, failure modes. Stress-test against the codebase before any code lands.
-
Split into tickets
Cut the plan into thin vertical slices — each one cuts through every layer end-to-end and is demoable alone. Tight feedback loops let the agent surface integration issues early and self-correct.
-
TDD red-green
Failing test first. Pass it with the minimum code. Refactor without changing behavior. The triple progression engine is one example — pure function, ~1,500 tests across the repo.
-
Push for review
PR with What/Why/How, pre-flight checks (lint, types, tests), reviewer assignment. Address tagged comments. Merge only when CI is green.
Read the source
Everything on this page lives in the repo.