Skills
A skill is a set of instructions for one kind of task (reviewing code, writing release notes, triaging bugs) that an agent loads when it needs it. Skills keep system prompts short: an agent carries a one-line description of each skill and reads the full instructions only for the task in front of it. Halo ships seven, and reads your own from a folder.
There are two ways to use them:
| Written | What happens |
|---|---|
skills: "code-review, house-style" in the agent's brackets, or skills: ["a", "b"] on a call | Those skills' instructions are added to the system prompt, every time |
tool Kit : Skills("./skills"), equipped, with tools: [Kit] on a call | The agent sees every skill's name and description, and loads the ones it needs with Kit.load |
tool Kit : Skills("./skills")
agent Notes [anthropic/claude-haiku-4-5, skills: "release-notes, house-style"]
agent Helper [anthropic/claude-sonnet-4-5]
equips Kit
pipeline Release(commits: String) -> Result:
commits |> Notes.write()
pipeline Anything(task: String) -> Result:
task |> Helper.work(tools: [Kit])skills: on a call replaces the agent's list for that call; skills: "none" turns them off. A skill that can't be found fails the call with UNKNOWN_SKILL, naming the skills there are.
Built-in skills
| Skill | For |
|---|---|
code-review | Finding bugs and security problems in a diff; answers JSON { verdict, summary, findings: [{ file, line, severity, issue, fix }] } |
write-tests | Adding focused tests in the project's existing framework and style |
web-research | Answering a question from sources, with a URL for every claim |
data-extraction | Pulling fields out of documents as strict JSON, null when absent |
release-notes | Turning commits or PRs into user-facing notes, breaking changes first |
incident-postmortem | A blameless postmortem: impact, timeline, root cause, action items |
bug-triage | Classifying an issue: kind, severity, duplicate, what's missing; answers JSON |
halo skills lists every skill a project can use (yours first), and halo skills code-review prints one. Several agent presets come with a built-in skill already loaded.
Writing a skill
A skill is a folder with a SKILL.md, the same layout as Agent Skills, so skills written for other agents work here too:
skills/
house-style/
SKILL.md
words.md # supporting files, read with Kit.read---
name: house-style
description: Write in our house style - British spelling, no jargon, sentence-case headings. Use for anything customers will read.
---
# House style
- British spelling (colour, organise).
- ...descriptionis what an agent reads to decide whether to load the skill, so say what it's for and when to use it.namedefaults to the folder's name. A single<name>.mdfile with the same frontmatter is a skill too.- Supporting files in the folder are listed in
load's result and read withKit.read(name, path); paths can't leave the folder, and symlinks out of it aren't followed. - A skill of yours with the same name as a built-in one replaces it.
The Skills tool
tool Kit : Skills("./skills") // your skills, then the built-in ones
tool Kit : Skills("./skills", builtin: false) // only yours
tool Kit : Skills() // ./skills, then the built-in ones| Method | Arguments | Returns |
|---|---|---|
list | [{ name, description, source }] | |
load | name | { name, description, instructions, files } |
read | name, path | { name, path, content } |
All three only read. When an agent is offered the tool, load's description lists every skill with its description, and its name argument is limited to those names. An agent that equips a Skills tool also looks up its skills: there; one that doesn't uses ./skills and the built-in skills.