mg-harness

mg-harness

Claude Code skills and harness rituals for engineering leaders.

The coding-agent ecosystem optimizes for engineers writing code. An engineering leader's job is wider than that — code, operational queues, communication, decisions, multi-week initiatives that span dozens of sessions. The skills and rituals here are the parts of that wider job I've shaped Claude around. They're small, opinionated, and meant to make your week less reactive, not just faster.

Quickstart (30-second setup)

/plugin marketplace add mgrabovskyi/skills
/plugin install mg-harness@mg-harness

That installs the plugin. Two surfaces show up:

• Skills — auto-triggered expertise. You describe what you want; the right skill loads. No commands to memorize.
• Harness — seven slash commands (/mg-harness:start-work, /mg-harness:grill, /mg-harness:checkpoint, …), six lifecycle hooks, and three subagents that turn a single session into a reliable multi-session workflow.

After install, set the two userConfig values for the verification hooks:

/plugin config mg-harness

Set verify_command (e.g. npm run lint --silent) and test_command (e.g. npm test --silent) for any repo you'll use the harness in. Both are optional; without them the harness still works, just without the auto-verify gates.

Why these skills and rituals exist

I built these to fix failure modes I kept hitting when applying a coding agent across the whole job, not just code.

#1 — The agent overbuilds and "improves" things you didn't ask it to touch

"Perfection is achieved not when there is nothing more to add, but when there is nothing left to take away."

— Antoine de Saint-Exupéry

The Problem. You ask for a small change. You get a sprawling diff. Adjacent code gets "improved." Variables get renamed. Error handling appears for impossible cases. Abstractions appear "for flexibility." Reviewing the PR takes longer than writing it would have.

The Fix is to load discipline rules before the agent writes anything:

• karpathy-coding-guidelines — surfaces assumptions, enforces minimal diffs, demands verifiable success criteria, and pushes back against speculative abstraction or unrequested cleanup.

The bar to clear: every changed line traces directly back to what you asked for.

Tip. Install this skill on the repo before the first agent-written change, not after. Once a codebase has been "improved" by an undisciplined agent, the cleanup cost is higher than the original work.

#2 — Triage is half mechanical work and half judgment, and the mechanical half eats your week

"There is nothing so useless as doing efficiently that which should not be done at all."

— Peter Drucker

The Problem. Most issue triage is mechanical: parse priority hints, route by domain, dedupe. Linear's native features (Triage Rules, Triage Intelligence) already cover the deterministic half. The judgment half — "this _Priority:_ p1 buried in the description means Urgent", "this is the same root cause as INF-2104" — is what actually burns calendar time when a human does it.

The Fix is to layer agent behavior on top of Linear's native triage, not in place of it:

• linear-triage-automation — runs when an issue enters Triage. Handles only the fuzzy parts: free-text priority parsing, duplicate confirmation, domain inference, and the decision to move out of Triage or leave for a human. Org-agnostic — takes a routing table and a fallback owner from the calling team.

If a triage task can be expressed deterministically, it does not belong in an agent loop.

#3 — Signal is scattered across Slack, Teams, and email, and the morning catch-up has no edges

"Clarity about what matters provides clarity about what does not."

— Cal Newport, Deep Work

The Problem. Your morning starts with hundreds of unread items across multiple tools. Most are noise. The useful items — explicit asks, places you were mentioned, things that shipped, things that are blocked — are buried.

The Fix is a scheduled task that does the synthesis before you sit down:

• morning-briefing — configures a recurring scheduled task. Each morning it pulls the last 24h from Slack/Teams + email and posts a four-section summary (needs from me, mentions, updates, blockers) to a destination you pick.

The skill builds and installs the task. The task does the daily work.

#4 — Handoffs between engineers lose context, and the new owner pays the tax

"Adding manpower to a late software project makes it later."

— Frederick P. Brooks Jr., The Mythical Man-Month