Skills

Skills are markdown playbooks that the agent loads on demand at the moment they're relevant. Where a plugin declares what to optimize, a Skill shapes how the agent thinks at a specific step — for example, how to draft research ideas or how to probe a problem from first principles.

Skill vs. Plugin in one line

A Skill sharpens how the agent reasons at one step (a markdown checklist loaded when it's relevant). A plugin describes what to optimize for a whole domain. They compose — use a Skill alone, a plugin alone, or both.

Do I need to write one?

No. Arbor ships with sensible default Skills that load automatically. Write your own only when you want to improve the agent's reasoning at a particular step.

Why Skills

LLM-driven research has predictable failure modes: skipping the thinking and jumping to plausible-sounding tweaks, reconstructing context from memory instead of reading state, proposing parameter changes instead of real mechanisms. A Skill is a concentrated dose of guidance that counteracts a specific failure mode — injected exactly when it matters, not buried in a giant system prompt.

The Skill format

A Skill is a markdown file with YAML front matter plus the instructions themselves:

---
name: idea_drafting
description: Structured idea-drafting workflow for IDEATE rounds.
when_to_apply: At the start of every IDEATE round, BEFORE drafting any candidate idea.
---

# SKILL: Idea Drafting

You are about to enter IDEATE. Read this once now. Apply every part before
you propose a single candidate...

Field	Purpose
name	Identifier used to register and reference the Skill.
description	One-line summary of what the Skill does.
when_to_apply	The trigger condition — when the agent should load and follow it.
body	The actual playbook the agent follows.

Bundled Skills

Arbor ships a small set of Skills out of the box, loaded automatically:

Skill	When it applies
idea_drafting	At the start of every IDEATE round, before drafting candidate ideas. Enforces the "mechanism, not knob" bar — real research directions over parameter tweaks.
first_principles_probe	When the agent should reason about a problem from first principles rather than pattern-matching to familiar solutions.

You can adjust which Skills are active for a single run from the intake chat — type / to use a slash command:

/skill load my_skill            # load one of your own Skills for this run
/skill unload first_principles_probe   # drop a default Skill this run
/skill reset                    # restore the defaults

Writing your own Skill

1. Create the folder .arbor/skills/ inside your project and add a markdown file there, e.g. <project>/.arbor/skills/my_skill.md. Arbor discovers project Skills from this folder; a project Skill with the same name overrides a bundled one.
2. Add the name, description, and when_to_apply front matter.
3. Write the playbook. Be concrete and prescriptive — a Skill is most effective when it gives the agent a checklist to apply, not vague encouragement.

Load it with /skill load my_skill in the chat (or rely on when_to_apply to trigger it automatically).

Skills vs. plugins

Reach for a plugin to define the eval contract, protected paths, and budgets for a domain. Reach for a Skill to improve the agent's reasoning at a particular step. They compose: a domain plugin can pair with Skills that sharpen ideation for that domain.