Harbor operator skills

Skills (in the Anthropic claude-code sense) for building agents with Harbor. Each one is a focused playbook: read it, follow the numbered steps, get a working surface.

These are operator skills — playbooks for humans (and AI coding agents) building Harbor agents. They are NOT the runtime skill subsystem the planner consults during reasoning (those live in internal/skills/). The glossary entry in docs/glossary.md pins the distinction.

The first-five-minutes adoption chain

A fresh contributor follows this exact chain to get from git clone to "I'm chatting with a working agent" in under five minutes:

1. scaffold-a-harbor-agent — harbor init + harbor scaffold + the tiered yaml.
2. run-the-dev-loop — harbor dev + harbor console + the token handshake.
3. drive-the-playground — chat against the agent, upload files, steer mid-run.

That's the load-bearing path. Everything else is built around it.

Skills by phase

Start a project

• scaffold-a-harbor-agent — start a fresh agent project from zero.
• define-the-agent-yaml — every field in harbor.yaml explained.

Build the agent

• add-an-in-process-tool — typed Go tools the planner can call in-process.
• wire-the-llm-provider — OpenRouter / Anthropic / OpenAI / Azure / NIM via Bifrost.
• configure-memory-and-skills — multi-turn memory + the runtime skill catalog.

Drive it interactively

• run-the-dev-loop — harbor dev + Console attach, single- or multi-process.
• drive-the-playground — chat, file uploads, multimodal dispatch, steering.

Observe + debug

• observe-with-the-console — a tour of the Console's Runtime / Execution / Resources / Settings pages.

Ship

• validate-and-package — harbor validate + make preflight + production checklist.

Build a custom frontend

• use-the-harbor-protocol — talk to the Runtime directly; ship a chatbot in a day.

Drift discipline

Per CLAUDE.md §18, any PR that changes a documented surface (a CLI verb, a Protocol method, a Console route, a harbor.yaml field) MUST update the matching skill in the same PR. Use this grep to find which skill is affected by a surface change:

grep -lE "surface: <changed-surface>" docs/skills/*/SKILL.md

The frontmatter audit (scripts/skills/check-frontmatter.sh, run as part of make drift-audit) catches shape drift. Content drift is human-reviewed; reviewers look for it.

Convention

Each skill is docs/skills/<slug>/SKILL.md with Dockyard-style frontmatter:

---
name: <slug-kebab-case>
description: "Use when <framing>"
license: Apache-2.0
metadata:
  framework: harbor
  surface: cli | agent-yaml | tools | mcp | llm | memory | playground | console | tasks | protocol
  verbs: "<harbor cli verbs>"
---

Sibling project Dockyard ships the same convention for MCP-server projects.