Skills: teach your agent a procedure once
You explain the same procedure to your agent for the fifth time: how to diagnose a crash, how to review a change for security, how to format answers for you. A skill writes it down once. The agent reads its one-line description at start, and loads the full procedure only when a task matches.
What you get. Procedures that apply themselves, at no standing cost: removing a skill and three sub-agents did not move our measured context floor by a single token. About thirty skills installed here, in daily use.
1. What a skill is
~/.claude/skills/diagnose-crash/
├── SKILL.md # name, description (the trigger), then the procedure
├── reporting.md # extra pages, read only if needed
└── scripts/ # optional helpers the procedure calls---
name: diagnose-crash
description: >
Diagnose why a program crashed on this machine, from a systemd-coredump
core dump. Use when a process segfaulted or aborted, or when asked why
an application crashed. Triggers: crash, segfault, core dump.
---
# Diagnosing a crash
Work from evidence. Start with `coredumpctl info <pid>` ...The description is what decides when the skill fires: say what it does and the words a request will contain. The body is the procedure, written for a colleague.
2. The ones we use
- Answers shaped for an ADHD reader: the next action first, numbered steps, the state restated each turn, one concrete thing to do at the end. Active in every session.
- Crash diagnosis from a core dump: facts first, resource exhaustion ruled out before blaming the program.
- Reverse engineering with evidence (see REA): analyze, prove, ask, then code.
- A security review harness: a router skill, a review pipeline, and one knowledge base per class of flaw (injection, XSS, SSRF, path traversal, secrets, race conditions…), loaded only for the class being hunted.
- Video as a web page (HyperFrames): an explainer, a promo or a captioned clip is written as an HTML composition with timing attributes, previewed in the browser and rendered to video. A family of skills covers animation, audio, keyframes and a registry of ready blocks; a media skill resolves music, sound effects, icons and voices into local files with a ledger of their sources.
- Diagrams: forty-five chart types as self-contained HTML with inline SVG, which first asks for your brand colours (see diagrams in code).
3. Skill or tool?
A procedure with several right answers (how to split a project, how to word a reply) is a skill: the model applies judgement. A check with one right answer (does this file validate, does this test pass) is a tool: a program that answers the same way every time. Do not ask a skill to deliver a verdict (see BMAD with small models).
Traps
A vague description never fires, and a greedy one fires everywhere. Write the triggers the way you actually phrase requests, and test with three real ones.
Read a community skill before installing it, like any program: it can tell the agent to run commands.
Sources
Found this useful? Tell us what to measure next.