Skills library

core-agent ships three Claude-Skills–shaped bundles in SKILLS/ at the repo root. They teach an agent how to configure and embed core-agent itself — the meta-use case is that core-agent can use them to help a user set up another core-agent.

Why ship these as skills

The reference docs (the Reference section) and the prescriptive guides (the Agent design section) cover the same material in static form. The skills cover the workflow form: when an agent is invoked with “help me set up core-agent for my project” or “configure an autonomous monitor for me,” the skill body is the runbook the agent walks through.

Same material, different surface. Pick what matches your read intent:

Reading first time, want depth? → Agent design (prescriptive) + Reference (schemas)
Already running an agent, want it to walk you through configuration? → install the skill below, ask the agent for help

The three bundles

`cli-setup`

Walks a user through configuring core-agent for interactive (TUI / REPL) use. Triggers on phrases like “set up core-agent”, “help me configure core-agent for my repo”, “what should my AGENTS.md look like”. Walks the four customization layers (provider → AGENTS.md → skills → tools) plus permission posture.

`autonomous-setup`

Walks a user through configuring an unattended core-agent — single-agent monitor or multi-agent team. Triggers on “set up an autonomous agent”, “run core-agent unattended”, “build a multi-agent team”. Includes patterns for single-agent monitors, multi-agent decomposition (parent + specialists), budget tuning, crash-resume.

`library-embedding`

Walks a Go developer through embedding core-agent in their own binary. Triggers on “how do I embed core-agent”, “use core-agent as a library”, “custom prompter”, “HTTP-served agent”. Covers the minimal embed, the seven extension points, and a full HTTP-served agent worked example.

Installing

The bundles are portable. To make them available to your core-agent instance, copy the bundle into a skills directory the agent loads.

Project-scoped install (just this repo):

mkdir -p .agents/skills
cp -r /path/to/core-agent/SKILLS/cli-setup .agents/skills/

After the copy, core-agent auto-discovers the skill on next launch. Verify with /skills in the TUI.

User-global install (every project for this operator):

mkdir -p ~/.core-agent/skills
cp -r /path/to/core-agent/SKILLS/cli-setup ~/.core-agent/skills/

core-agent (v2.1+) auto-discovers both ~/.core-agent/skills/ (user-global) and .agents/skills/ (project-scoped) and merges them at load time. On name collision, the project-scoped skill wins — useful for forking the generic bundle and shipping a project-specific variant under the same name.

Using a skill once installed

Trigger the skill by phrasing your request to match its description:

$ core-agent
> help me set up core-agent for my repo
[agent invokes cli-setup skill, walks the four layers, writes config files as you go]

The agent reads the skill’s SKILL.md, follows the numbered procedure, and fetches references/*.md files on demand for the specific topic the user asks about.

To inspect what’s loaded:

> /skills
[lists all loaded skills with their descriptions]

To verify a skill triggered (or diagnose why it didn’t):

> /btw why didn't you invoke the cli-setup skill for the prior turn?
[the model often explains: "the request didn't match the description's phrasing" — and that's your signal to refine the description if it's your own skill]

Adapting

These bundles teach core-agent’s defaults and the patterns from the Agent design docs. If your organization has stricter standards (e.g., “every project must use anthropic-vertex with mode=ask”), fork the bundle and modify the body or references accordingly. The skill format is just markdown + YAML frontmatter; everything is editable.

For writing your own skills from scratch (not derived from these bundles), see:

Agent design → Skills — design patterns
Reference → Skills — SKILL.md schema, discovery, permission gating
Claude Skills spec — upstream format spec we mirror

Roadmap

Done in v2.1: ~/.core-agent/skills/ auto-discovery. The merge picks project-scoped skills over user-global ones on name collision, so you can fork a bundled skill into your project and the project version takes effect transparently.
v2.1+: First-run bundle install. A core-agent --install-skills (or auto-install on first run if ~/.core-agent/skills/ is empty) so operators don’t have to know about the cp step at all.
Future: more bundles. Specific role-shaped skills (code-reviewer-setup, devops-monitor-setup, etc.) once the three above prove out.