Specimens, carefully kept. Herbarium Claudii — a folio for living knowledge.
The herbarium is a workshop's memory. A pressed leaf, a date, a hand-written label — and decades later the species, the season, and the soil are all still legible. Skills are that, for Claude Code: small folders of preserved expertise, ready the moment the task they describe walks into the room.
Coll. — coastal scree, May.
Det. C. Code
The feature today is Skills.
A Skill is a folder. Inside it lives a SKILL.md — a short description
of when to use it — and any reference docs, scripts, or templates the task needs. When a
user's request matches the description, Claude Code pulls just that specimen off the shelf,
reads it, and works with the procedure already in hand.
In folklore sage is the keeper-herb — planted by the door, hung in the rafters, brewed when memory needs to stay sharp. Skills do the same job for a workshop of code: a tiny dossier that says, in plain prose, here is how we do this around here. The model doesn't have to remember; it has only to recognise.
Anatomy of a skill.
Three parts to a well-pressed specimen.
Every skill, however elaborate inside, presents the same outer label: a name, a description, and the body of preserved instructions. The description is the species tag — it's how the right specimen gets pulled when the right request arrives.
| Part | Path | Purpose |
|---|---|---|
| Cover label | .claude/skills/<name>/SKILL.md | Triggers — what tasks invite this skill. |
| Reference matter | .claude/skills/<name>/*.md | Deeper notes — loaded only when needed. |
| Pressed instruments | .claude/skills/<name>/scripts/ | Optional helpers the skill can run. |
Claude Code reads the description for every skill at session start, but the body
of each SKILL.md is only opened when its specimen is needed. The shelf can grow
to dozens of folders without crowding the conversation.
Field observation.
Naming the leaf you wish to press.
A new skill is a folder and a markdown file. The frontmatter teaches Claude Code when to reach for it; the body teaches it what to do once it has. Here, a small one for writing blameless postmortems:
$ mkdir -p .claude/skills/postmortem $ cat > .claude/skills/postmortem/SKILL.md <<'EOF' --- name: postmortem description: Write a blameless postmortem from an incident timeline. Trigger on "write a postmortem", a Pager URL, or when the user shares an outage summary after resolution. --- # Postmortem — house style 1. Lead with impact (minutes, users, $). 2. Timeline in UTC, one event per line. 3. Root cause separated from contributing factors. 4. Action items have owners and dates. EOF $ claude › write a postmortem from incident-2026-05-13.md ↳ Using skill: postmortem (.claude/skills/postmortem/)
That's the whole ritual. Once the leaf is mounted, the next time someone asks for a postmortem, the right method is right there — no copy-and-paste from the last incident, no reminder needed about the house style.
The herbarium is a library of evidence — every sheet a fact you can return to. — Anne Pratt, after the manner of