Skills For Real Non-Engineers

I've led dev teams for over twenty years. I never wrote the code. These skills give me a lead developer who runs the team, so all I have to bring is the idea. Four shipped products so far.

Most of these skills started as someone else's. Fifteen are adapted from Matt Pocock's Skills For Real Engineers, three from Lauren Tan, six are mine. I rewrote nearly all of them for someone who won't read the code, then tied them together under one orchestrator session that takes the idea, decides what gets grilled, mapped, prototyped, built, and reviewed, and runs the rest. That's what lets a non-engineer ship at production quality. I talk to that one session the way I've always talked to a dev lead.

Each skill is a SKILL.md, a plain-markdown playbook the agent reads on its own: every workflow I used to re-explain each session, written down once, installed in one line and versioned like software.

If you just want to install something, jump to Installation. If you want to know why any of this exists, read on; it's short.

Why these skills exist

Every skill here started the same way: an agent failed me in a pattern, twice, and I got tired of re-explaining the fix. These are the four failures that cost me the most.

The agent built the wrong thing

Nobody describes what they want completely, me included. I'd ask for a change, the agent would fill every unstated gap with a guess, and I'd find the guesses three files deep in the diff.

The fix is grilling: the agent interviews me about the plan, in rounds, until nothing load-bearing is still assumed. Facts are the agent's job to go fetch; decisions stay mine. I run it before any change big enough to argue about, which turns out to be most of them.

A hard decision got one model's opinion

Some calls are expensive to get wrong: an architecture, a contract, a launch plan. Settling one with whichever model happened to be in the chair that session never sat right with me.

The fix is advisory-board: Claude, Codex, Gemini, and Grok each review the same decision independently, read each other's notes, argue out the disagreements, and hand back one recommendation you read like a memo. It's read-only by default and you approve exactly what leaves your machine, or you run a fully local board. Point it at a codebase and the advisors cite path:line evidence; every run also emits a machine-readable verdict.json you can gate CI on. There's a full tour with real example runs, and the runs themselves are in examples/.

Everything the agent writes sounds like AI

The em dashes. The "not just X, but Y". The victory lap at the end of every reply. Read enough agent prose in a day and the tells start to grate.

The fix is plainspoken, which is always on once installed: plain words, concrete claims, no tells, in everything the agent emits, commit messages included. For pages a stranger will read, writing-for-humans adds what clean prose alone can't: structure, warmth, an admitted limit or two. And when a message has already failed to land, huh translates it.