Skills

A skill is a reusable instruction bundle — a SKILL.md with a name, a short description, and a body of guidance the model should follow when the task calls for it. Skills install from the internet and their instructions are injected into agent and primary-runtime context.

Use a skill when you want to teach the workspace how to do something (fill PDF forms, write release notes a certain way, follow a house code style) without repeating the instructions in every message.

Install from the internet

In the right-rail Skills pane, Install from the internet fetches a SKILL.md and stores the skill in a per-workspace catalog (.hive/skills.json). The same reference forms as MCP work:

https://raw.githubusercontent.com/owner/repo/main/pdf/SKILL.md
https://github.com/owner/repo/blob/main/pdf/SKILL.md   # blob URLs become raw
owner/repo/pdf/SKILL.md                                # shorthand, ref defaults to main
https://skills.sh/some/skill                           # any https host

SKILL.md format

YAML-ish frontmatter (all optional) followed by the instruction body:

---
name: PDF Filler
description: Fill PDF forms from structured data
allowed-tools: Read, Write, Bash
---
# PDF Filler

Use `pdftk` to fill forms. Always validate the output with
`pdftk … dump_data_fields` before returning.

• name → the skill's display name (falls back to the folder name).
• Everything after the frontmatter is the instructions body.
• Other frontmatter keys (description, allowed-tools) are advisory and aren't stored on the skill — Hive keeps only the name and the instruction body (plus the source URL).

No frontmatter is fine — the whole file becomes the instructions and the URL's folder name is used as the skill name.

How skills reach the model

Skills loaded into a chat are injected in full into the participants' system prompt. They appear under a shared heading:

Loaded skills (follow these):

## PDF Filler
<the skill's instruction body>

## Some Other Skill
<its instruction body>

Each skill is a ## <name> block. There is no autoload toggle and no per-scope setting — every skill loaded into the chat is injected in full; the model follows the ones relevant to the task at hand. Remove a skill from the Skills pane if you don't want it in context.

Targeting specific agents

By default a skill is global — it applies to the primary and every agent (the behavior above). A skill can instead be targeted at specific agents, so only those agents get its instructions and the primary never receives it.

The Skills pane's add-form has an agent-target picker — a row of agent checkboxes. Leave all unchecked for a global skill (all agents); check one or more to scope the skill to just those agents. Each skill row shows a scope badge (global, or the targeted agents).

Utilization is otherwise unchanged: a skill's instructions are injected in full into the system prompt of whichever participants it targets.

Where it lives

Installed skills are workspace-level: stored in .hive/skills.json, seeded into every chat in the workspace, and surviving relaunch. Remove a skill from the Skills pane to drop it from the catalog.

Public skills only, for now

Skill fetches are unauthenticated, so private-repo SKILL.md files aren't supported yet. Host the skill at a public raw URL or via skills.sh.