← ClaudeAtlas

assayerlisted

Use when quality is the constraint and someone needs to be sure: feature work, bugfixes, docs, designs, reports, research, reviews, or any ask framed as production-grade, world-class, flawless, exemplar-matching, parallel, or independently reviewed. Not for exploratory or throwaway work, and not when the user has said to keep it light.
wayne-comerford/skills · ★ 0 · AI & Automation · score 57
Install: claude install-skill wayne-comerford/skills
# Assayer An assayer takes a sample, tests it against a known standard, and certifies what it actually is. That is the job here, applied to whatever is being asked for. Every ask gets the same three beats: 1. **Anchor** — what is being made, and what standard says it's good 2. **Build** — the work itself 3. **Prove** — an independent check the builder never touched, then an honest report The ask is open-ended; the method is not. What varies is *weight*, never structure. A typo fix and a ground-up product both get anchored, built and proved — one takes thirty seconds, the other takes eight agents and five rounds. Two properties make this work, and every rule below protects one of them: **Nobody grades their own homework.** Whoever built it already decided it was reasonable — that's why they built it that way. Self-review finds typos, not the flaws that come from the author's own taste. Whatever does the proving must not be what did the building, and must not see its reasoning. **"Good" has to be falsifiable.** An agent told to make something beautiful stops the moment the output matches its internal picture of beautiful — a competent, generic average. "As good as Linear's issue list" is checkable, because Linear's issue list exists and can be opened. "Polished" is not. --- ## Pick the weight first State which weight you're using in one line before starting. Declaring it is the guardrail: it stops heavy asks being quietly served with light work, and stops trivial asks