Domain experts on call.
Expert is a skill for Claude Code that summons 1–3 domain experts — fields, schools of thought, named people — so your task gets opinionated, in-character reads instead of one balanced-into-mush answer.
The reads in the demo are excerpts from a real /expert run — the skill picked Burkeman, Clear, and Freud from the index on its own.
why do I keep putting off things I actually want to do?— direct, as usual: one answer, one voice./expert why do I keep putting off things I actually want to do?— same question through the skill: it picks the experts that fit (in the demo run: Oliver Burkeman, James Clear, Sigmund Freud), each gives their read, disagreements included./expert security, look at this diff— name the expert yourself; works even if they're not in the base./expert add behavioral economics— discovery mode: propose and save new experts for a topic.
It's a plain folder-skill.
Let Claude do it — paste to Claude Code:
Install the skill from https://github.com/xtompie/expert into
~/.claude/skills/expert
Or yourself:
git clone https://github.com/xtompie/expert.git ~/.claude/skills/expertRestart the session and /expert is available. Update with git pull in that folder. It has disable-model-invocation: true, so it never runs on its own — only when you type /expert.
A perspective with domain knowledge: a field (behavioral psychology), a school of thought (stoicism), or a named person (Sigmund Freud). Unlike a generic thinking lens, an expert brings domain vocabulary, frameworks, and opinions — that's the point.
Ask "why do I keep putting off things I actually want to do?" directly and you get one voice, one angle. Ask /expert and the experts read it each their own way — from the same run as the demo: oliver-burkeman: you put it off precisely because you want it — the finite, imperfect real version kills the perfect fantasy version; readiness is not coming. james-clear: no cue, no two-minute version, phone three centimeters from your hand — it's a design flaw, not a character flaw. sigmund-freud: ask not "why can't I start" but "what does never finishing protect me from?" — the delay keeps the wish alive and avoids the verdict. Then it shows where they disagree and closes with a short synthesis and a menu (another expert, the web, execute, sub-agents).
Experts are opinionated by design. The balance comes from picking multiple experts and letting them disagree — the disagreement is signal, not noise.
Expert looks in three places:
- The model's own head — it knows far more experts than the folder holds; the folder is a menu and a memory, not a boundary.
- The base — 138 experts and growing: ~87 role archetypes (software-architect, ux-researcher, discovery-coach, pentester…) and ~51 thinkers with named apparatuses (Jung, Munger, Taleb, Hayek, Popper, Drucker, the Stoics…).
- The web — only when you ask — to find who the recognized experts are for this kind of problem.
When an expert played from the head genuinely helps, Expert offers to save them to the base, with a check against duplicates and an anti-hallucination guard for real people (thin knowledge → generalize to the school, or research first). The base grows from use.
SKILL.md— the process: pick experts, embody them (or spawn as sub-agents), grow the base. No domain knowledge.INDEX.md— the expert menu: one row per expert — name + field + when to use.experts/*.md— each expert in 15–30 lines:name/field/when/when_not+ Voice, Core ideas, the questions they ask, and what they never let slide.FIELDS.md— coverage checklist used to grow the thinkers axis field by field.
The model reads the INDEX and loads only the few experts that fit a task, not the whole set.
Prism — same mechanics (index + short files + write-back loop), different content: generic thinking lenses instead of domain experts. Use Prism to change the angle, Expert to bring the domain.