Engineering Leader Skills
Claude Code skills for engineering managers and tech leads, by Matt Kaszubski.
v0.2.1, small and opinionated by design. The triggers and rubrics sharpen as they meet real PRs and real teams; that iteration is the point, not a caveat. Issues welcome, especially where a skill misfired.
What this is
A small, curated collection of Claude Code skills for the work of leading engineers: reviewing code, coaching people, reading a team's health. Not a grab-bag.
Most skills are anchored on an established leadership framework, operationalised so the framework actually changes the output. One of them, repo-xray, is the collection's measurement instrument: not framework-anchored, just careful, calibrated measurement. Either way, the collection shares one vocabulary.
The bet: plenty of general-purpose Claude skills exist; almost none are built deliberately for engineering leaders. This fills that gap.
The skills
| Skill | What it does | Anchored on |
|---|
repo-xray | Reads git + PR history and surfaces non-obvious team-health signals, narrated. Local, deterministic, advisory. | Measurement instrument (DDD lens, one signal) |
pr-review | Reviews a PR for what scales or sinks a team (coupling, reviewability, convention drift), not style nits. | Radical Candor |
coaching-calibrator | Given an engineer and a task, recommends how to coach or delegate it. | Situational Leadership II |
retro-facilitator | Turns raw retro material into a facilitated retro: separates fact from story, converts blame into ownership. | Conscious Leadership |
Examples
repo-xray
"Run a repo-health x-ray on this project — last 365 days"
Runs five signal queries (review concentration, knowledge silos, time-to-first-review, stale PRs, silent merges), then narrates them as a calibrated health note. Calibrated meaning: a 66% silent-merge rate on a two-person repo is not a 66% silent-merge rate on a thirty-person team, and the narration says so. Runs entirely on your machine with your own git/gh auth; nothing is posted.
pr-review
"Review PR #42 on owner/repo"
Returns a Radical Candor review: Verdict / Blocking / Non-blocking / Done well / Self-check. Cites file:line. Never posts to GitHub; advisory only.
Sample output (real review of a small UX PR)
**Verdict:** Clean batch. Nothing blocking — one bundling note and two polish items.
**Non-blocking** — worth considering, reviewer's call
- PR bundles three unrelated concerns. Marquee toggle, justify sweep, and arrow
glyph swap have nothing in common except "UX walkthrough output." Small enough
to hold in your head here, but the pattern degrades fast if a future "Batch N"
puts a 50-line behavior change next to silent CSS edits.
- `e2e/marquee-toggle.spec.ts:50` — fixed `waitForTimeout(700)` after the pause
click. You used `expect.poll` earlier in the same test. Mirror that pattern —
the brittle version will flake under CI load.
**Done well**
- The pause test asserts behavior (`.worked__track-shift` transform stops
advancing), not just `aria-pressed`. That's the test that catches a real
regression.
**Self-check:** Passes Radical Candor — each item names file/line and a concrete change.
coaching-calibrator
"Sarah is moving onto incident response. She handled one ticket before with help, isn't confident yet. How should I support her?"
Diagnoses competence and commitment separately, lands on a development level (D1–D4), then prescribes the matching style (S1 Directing → S4 Delegating) as concrete behaviours for the week. Asks for missing signals rather than guessing a level, because SLII is a conversation, not a covert label.
retro-facilitator
"Run a retro on Friday's outage — the deploy went out at 4pm and on-call spent the weekend on it."
Structures the debrief through Conscious Leadership. It reads the room's state (above or below the line), splits every observation into a verifiable fact and a labelled story, turns victim/villain/hero framing into ownership spread across the people involved, then turns what they own into concrete experiments for next time. Here's the move a generic "what went well / badly" template skips: it won't let a story ride in as a fact, and it names no villains.
Foundations
The skills draw on these frameworks, distilled for application in foundations/:
These are distillations for use, not reproductions; each one credits and links its source.
Install
As a plugin (recommended)
In Claude Code:
