Claude Code skills

Use this to remember which skills exist, which ones we're not firing, and how to write our own. Verified 2026-08-23 against the plugin cache, marketplace registry and skills directories on Teagan's laptop. A designed version of this page is published as a private artifact — see the link at the bottom.

What a skill actually is

A folder with a SKILL.md in it. The frontmatter carries a name and a description; the body carries instructions. That is the whole format.

The mechanism worth understanding is when the body gets read. Every skill's description sits in context permanently — a few dozen tokens, always there, acting as a trigger. The instructions underneath only load when something matches. A rack of forty skills costs you forty descriptions, not forty manuals. That is why skills scale where stuffing everything into CLAUDE.md does not.

Two things fire a skill: the model matches your request against the description and loads it on its own, or you type /name and force it. Most built-in skills are written to fire automatically — you have probably been using artifact-design and dataviz for months without knowing their names.

This matters for our bill. Subagent fan-out is our largest controllable overspend — the finding behind XDEC-103 and the reason the subagent-discipline rules exist. A lot of what we reach for a subagent to do is a skill's job: encoding a procedure we keep re-explaining. Reach for the skill first. It runs in the context you already paid for.

The rack — what's installed

Everything below is available right now, in your seat container and on your laptop both. Underused marks the ones we have and aren't firing.

Shipping code

Things people look at

Automation and the harness

Reference and lookup

The shelf we haven't touched

The official marketplace is already registered on this machine. Only the Stripe plugin has ever been installed from it. These map onto problems we actually have.

Also on the shelf: pr-review-toolkit, commit-commands, code-modernization, frontend-design, plugin-dev, mcp-server-dev, agent-sdk-dev, and language servers for TypeScript, Python, Ruby, Rust, Go, C# and Swift.

Installing

The marketplace is registered already, so this is one command per plugin. Restart Claude Code afterwards.

# see what a plugin adds, and what it costs you in context
claude plugin details skill-creator@claude-plugins-official

# install
claude plugin install skill-creator@claude-plugins-official
claude plugin install session-report@claude-plugins-official
claude plugin install claude-md-management@claude-plugins-official

# check what you have
claude plugin list

claude plugin details before install is the habit worth keeping — it prints the component inventory and the projected token cost, so you can see what a plugin does to your context budget before it does it.

Nobody here has written one

This is the finding. There is no ~/.claude/skills/ on this machine and no .claude/skills/ in any repo under C:\dev. Zero custom skills, across a fleet with a documented list of procedures we re-explain constantly.

Every one of these has already cost us something: the Bitwarden CLI quirk that silently burns a session, the duplicate TTS service deployed because nobody checked the capability registry, the access grant recorded in one project's memory store and therefore invisible to the seat agent that refused Lew. In each case the knowledge existed in dev-context and was not read at the moment it mattered.

That is the gap a skill closes. A document has to be found. A skill fires on the words you were already typing.

1. Decide where it lives

~/.claude/skills/<name>/SKILL.md is personal — your machine only, invisible to everyone else. <repo>/.claude/skills/<name>/SKILL.md is committed, so every seat container that clones the repo gets it.

For a three-person team, commit it. A procedure living on one laptop is the same failure mode as a grant living in one memory store.

2. Write the description as a trigger, not a summary

The description is the only part always in context, and it is doing one job: matching. Write the literal words someone would use — verbs, nouns, error strings, tool names. The shipped Stripe skills have descriptions running six lines of comma-separated triggers; that is not sloppiness, it is the format working correctly.

"Helps with secrets" never fires. "Use when retrieving a credential, API token, or password from Bitwarden, when bw get or bw list returns empty, or when a session needs a key from the SPS Service collection" fires every time.

3. Keep the body short and link out

Instructions in SKILL.md, longer references in sibling files the body points at. Same progressive-disclosure logic one level down.

---
name: bw-secret
description: Use when retrieving any credential, API token, password, or
  key from Bitwarden — including when `bw get` or `bw list` returns empty,
  when a command needs a secret from the SPS Service collection, or when
  BW_SESSION appears not to work. Covers the pipe-stdin workaround for
  bw 2026.4.1 and the custom-fields retrieval path.
allowed-tools: Bash, Read
---

# Retrieving a secret from Bitwarden

Bitwarden is already unlocked in every environment. Never ask Teagan,
Lew or Raghav to unlock it.

## The 2026.4.1 gotcha
`list` and `get` ignore both `BW_SESSION` and `--session`. Pipe the
master password to stdin instead. `bw status` reports `locked` even
after a successful unlock — ignore it.

## Keys live in custom fields
Use `bw get item <name>` and read `fields[]`. Do not assume the
credential is in the password field.

Full reference: ~/dev-context/bitwarden-cli-usage.md

name and description are the required pair. allowed-tools narrows what the skill may touch; disable-model-invocation: true makes it slash-command-only when you don't want it firing on its own.

Six worth writing first

Each is a procedure already written down somewhere in dev-context and already skipped at least once.

The natural home for the shared ones is a committed .claude/skills/ in Workstation Hub, which all three of us already have checked out.

House habits

• Skill before subagent. If the answer is "a procedure we keep re-explaining," that's a skill and it's nearly free. Spawn an agent when the work genuinely needs its own context window — a broad search, real parallelism — not to carry instructions.
• Descriptions are the surface. A skill that never fires is worse than no skill, because you think it's covered. If one isn't triggering, the description is wrong — add the literal phrases you actually type.
• Commit them. Personal skills solve your problem once. A skill in a repo's .claude/skills/ reaches every seat container the moment they pull.
• Don't skill a one-off. Third time you explain the same thing, write it down. Not the first.
• Check the cost. claude plugin details <name> shows projected token cost before you install. A plugin you never use still spends its descriptions on every turn.
• Say the number. Skills carry defaults you may not want — /loop reaches for 300–1200s when our cadence is 270s daytime and 60s overnight. State the constraint rather than accepting the default.

Caveats

Verified against Teagan's laptop install on 2026-08-23 — not against Lew's or Raghav's seat containers, which may ship a different set if their Claude Code build is older. Run claude plugin list in your own environment to confirm.

The canonical, always-current sources are ~/dev-context/ and the intranet. Where this page and those disagree, those win.