Skill
Also: 技能 · agent skill · 技能包 · SKILL.md
A packaged, reusable capability: an instruction file for one class of task, plus optional scripts and resources, loaded by the agent only when needed.
When you will meet it
You reach for it when you want the agent to follow your own way of doing a task instead of re-explaining every time. The blocker is almost never authoring — anyone can write a Markdown file; it is installing two hundred skills and watching zero trigger. Discovery and triggering are the hard part, not writing.
An analogy
Like a shelf of SOP binders: nobody memorises them, but everyone must know which binder to pull for which job. The best-written binder that nobody can recognise on the shelf might as well not exist.
Minimal example
skills/hk-stock-dd/
SKILL.md ← 何時用我（description）＋ 怎麼做（正文）
scripts/ ← 選配：可執行腳本
references/ ← 選配：模板、檢查清單
載入方式（示意）：
平常：上下文裡只有每個技能的一行 description（省 token）
命中：Agent 判斷任務相關 → 讀入整份 SKILL.md → 照著做Note the two stages: the description decides whether the skill is ever thought of; the body decides how well it runs. Most failures happen at stage one — a description missing the words users actually say (especially non-English ones) means the skill never triggers.
What people get wrong
- Assuming authoring is the hard part. Triggering is: the agent matches the description against the user's words — you wrote in your jargon, users speak in theirs, and a beautifully written skill never loads.
- Stuffing skills into the system prompt as always-on text. Hundreds of full skills devour the context window; the essence of a skill is lazy loading — only its one-line summary is present by default.