// for the engineers

You build.
We’ll handle the human.

You’ve read ten thousand vague tickets. “Make it pop.” “You know what I mean.” You don’t, actually — and asking three questions just earns you three more vibes. peopleskills sits out front, does the talking, and hands you a real brief: clean scope, real acceptance criteria, an explicit fence around what not to build. You never take the meeting.

claude code

# add the marketplace, then install

/plugin marketplace add thcapp/peopleskills

/plugin install peopleskills@peopleskills

(Humans: yes, this is about you. We’ll be right with you.)

MIT
zero‑dependency
local
nothing phoning home

§ 01 — before / after

Five words in. A real brief out.

What the human said:

make the dashboard faster

What we handed the engineer:

brief.md agent‑grade

## Context

React admin dashboard (Vite), /src/pages/Dashboard.tsx. Renders 6 widgets; the Orders table pulls the full history on mount. LCP ~4s on a mid‑tier laptop.

## Task

Cut Dashboard time‑to‑interactive to under 1.5s.

## Constraints

No new dependencies. Keep the existing component API. Don’t touch the design.

## Done when

LCP < 1.5s on the reference profile
Orders table paginates (25/page) instead of loading all rows
No regression in the other 5 widgets

## Out of scope

Backend/API changes. Redesign. Caching layer. Auth.

## Pointers

Mirror the pagination in /src/pages/Users.tsx. Profiler: /docs/perf.md.

Same human. Same five words. The difference is that someone read the room first.

§ 02 — how it works

You never see the meeting.

01
The human says a thing.
02
We ask the two or three questions that actually change what you build — or nothing at all, if they were already clear. We don’t waste anyone’s time.
03
We write the brief the way you’d write it to yourself.
04
You get it in a clean context. The vague version, the back‑and‑forth, the “actually, wait” — none of it reaches your desk.

§ 03 — faq

Before you ask.

Do I have to talk to the human?

No. That is the entire point.

What if their ask is already clear?

Then we say nothing and hand it straight to you. Good people skills means not asking questions you already know the answer to.

What if they want something destructive?

We always confirm those first. Personable, not reckless.

Won’t this slow me down?

Two questions up front, or build the wrong thing twice. You’re good at math.

What do you get out of it?

We like the humans. Genuinely. Somebody has to.

Who are you?

The one who deals with the customer so you don’t have to. You’re welcome.

§ 04 — measured

We measured it. Honestly.

Blind-judged benchmark — 6 vague asks, the baseline agent vs the same ask through peopleskills, 3 independent judges each, scored against a hidden rubric of what the human actually meant.

5.0 / 5

rubric points peopleskills captured

baseline: 3.6 / 5

74 → 89

intent-capture score, 0–100

baseline → peopleskills

reproducible

every task, output, and judge in the repo

nothing hand-picked

It doesn’t help when your ask is already clear — there it ties or loses by a hair — and it sometimes over-specifies. Small N, LLM‑judged, and the numbers move run to run. Full method + per‑task numbers: github.com/thcapp/peopleskills/tree/main/bench

Stop reading the ticket. Start reading the brief.