Writing Hive SkillsSAFE

Multi-Agent Harness for Production AI

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: hive.writing-hive-skills
description: Author a new Agent Skill for a Hive agent that conforms to the Agent Skills specification (SKILL.md with YAML frontmatter, optional scripts/references/assets directories). Use when the user asks to create, scaffold, add, or package a new skill for a Hive agent.
metadata:
  author: hive
  type: default-skill
  spec-source: https://agentskills.io/specification
---

## Operational Protocol: Writing Hive Skills

Hive agents discover skills by scanning several roots, in precedence order:

1. `<project>/.hive/skills/` — project, Hive-specific
2. `<project>/.agents/skills/` — project, cross-client
3. `~/.hive/skills/` — user, Hive-specific
4. `~/.agents/skills/` — user, cross-client
5. Framework defaults shipped in `core/framework/skills/_default_skills/`

Each skill is a directory containing a `SKILL.md`. At startup, only the frontmatter `name` + `description` of every skill is loaded; the body is loaded only when the agent activates the skill. Design for that.

### Choosing where to put a new skill

- **Colony-scoped (via `write_skill` inside colony mode)**: when the skill is the operational protocol a single colony needs — its API auth, DOM selectors, DB schema, task-queue conventions — do NOT place it under `~/.hive/skills/` or `<project>/.hive/skills/` yourself. Those roots are SHARED and every colony on the machine will see it. Instead, once you're inside the colony (post-fork), call `write_skill(skill_name, skill_description, skill_body, skill_files?)` and the skill is materialized under `~/.hive/colonies/<colony_name>/skills/<skill-name>/` where only that colony's workers discover it. See the subsection below.
- **Project-scoped**: put under `<project>/.hive/skills/` when the skill is tied to that codebase's APIs, conventions, or infra and multiple agents in the project should share it.
- **User-scoped**: put under `~/.hive/skills/` when the skill is reusable across projects for this machine/user and all agents should see it.
- **Framework default**: add under `core/framework/skills/_default_skills/` AND register in `framework/skills/defaults.py::SKILL_REGISTRY` only when the skill is a universal operational protocol shipped with Hive. Default skills use the `hive.<name>` naming convention and include `type: default-skill` in metadata.

### Colony-scoped skills via `write_skill`

A colony-scoped skill is one that belongs to exactly ONE colony — e.g. it encodes the HoneyComb staging API the `honeycomb_research` colony polls, or the LinkedIn outbound flow the `linkedin_outbound_campaign` colony runs. Writing such a skill at `~/.hive/skills/` or `<project>/.hive/skills/` leaks it to every other colony, which will then see it at selection time.

**Do not create the folder yourself with the terminal.** Once the colony has been forked (via `suggest_colony` → user confirms the Create Colony popup), call `write_skill` inside the colony:

```
write_skill(
    skill_name="honeycomb-api-protocol",
    skill_description="How to query the HoneyComb staging API...",
    skill_body="## Operational Protocol\n\nAuth: ...",
    skill_files=[{"path": "scripts/fetch_tickers.py", "content": "..."}],  # optional
)
```

The tool writes `~/.hive/colonies/<this_colony>/skills/honeycomb-api-protocol/SKILL.md` (plus any `skill_files`), which `SkillDiscovery` picks up as project scope when that colony's workers start — and ONLY that colony's workers. No cross-colony leakage.

Do not write colony-bound skill folders by hand under `~/.hive/skills/`. A skill placed there is user-scoped and becomes visible to every colony on the machine — defeating the isolation you wanted.

### Directory layout

```
<skill-name>/
├── SKILL.md          # Required
├── scripts/          # Optional — executable helpers
├── references/       # Optional — on-demand docs
└── assets/           # Optional — templates, data, images
```

Rules:
- The directory name **must** equal the `name` frontmatter field (for framework defaults, the directory is the unprefixed name, e.g. `writing-hive-skills/` for `hive.writing-hive-skills`).
- Keep `SKILL.md` under ~500 lines. Move long reference material into `references/`.
- Reference other files with relative paths from the skill root (`scripts/foo.py`, `references/API.md`). Keep references one level deep.

### SKILL.md frontmatter

Required fields:

| Field | Constraints |
|-------|-------------|
| `name` | 1–64 chars, `[a-z0-9-]`, no leading/trailing/consecutive hyphens. Must match the directory name. Framework defaults prefix with `hive.` |
| `description` | 1–1024 chars. Must describe **what** the skill does **and when to use it**. Include trigger keywords the user is likely to say. |

Optional fields:

| Field | Notes |
|-------|-------|
| `license` | License name or reference to a bundled file |
| `compatibility` | ≤500 chars. Only include if env requirements are non-trivial (network, tools, runtime) |
| `metadata` | Free-form string→string map. Namespace keys to avoid collisions. Default skills set `type: default-skill`. |
| `allowed-tools` | Experimental. Space-separated pre-approved tools, e.g. `Bash(curl:*) Bash(jq:*) Read` |

Minimal template:

```markdown
---
name: my-skill
description: One sentence on what it does. One sentence on when to use it, with concrete trigger words the agent will see in user requests.
---

# My Skill

<body>
```

### Writing a good `description`

This is the single most important field — it's the only thing the agent sees at skill-selection time.

- **Bad**: `Helps with trading.`
- **Good**: `Buy and sell shares on the HoneyComb exchange. Handles auth, slippage-protected orders, idempotent retries, and AMM output estimation. Use when placing trades or interacting with the AMM.`

Include verbs the user is likely to say (`buy`, `sell`, `place trade`) and proper nouns (`HoneyComb`, `AMM`).

### Writing the body

Structure the body for the agent, not a human reader:

1. **Lead with what the agent can't guess** — API base 

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (1)

CLAUDE.md

Gates applied: no_behavioural_pass.