# Skills, Rules, and Context (/web/agents/skills-rules-context)



Mogplex splits shared agent building blocks into three different libraries on
purpose:

* skills
* rules
* context

They are related, but they are not interchangeable.

## Skills [#skills]

Skills are reusable capability packages.

Use them when an agent should know **how to do a class of work**, such as:

* following a specialist workflow
* using a reusable investigation pattern
* applying a repeatable framework-specific playbook
* importing an external skill from a registry

In the current product, the Skills surface has three acquisition paths:

* **local skills** you create and edit directly
* **skills.sh registry** browsing and install
* **Vercel docs** import, which can turn a Vercel document into a local skill

Local skills are typed as:

* runbook
* tool
* prompt
* workflow

That makes the Skills area both a library editor and a discovery surface.

## Rules [#rules]

Rules are shared instruction blocks and policy constraints.

Use them when an agent should consistently respect **how it is allowed to
operate**, such as:

* coding standards
* repo guardrails
* deployment restrictions
* review style
* behavioral policy

The current Rules surface is a small global file manager. You can create, edit,
and delete shared rule files, then reuse them across many agents instead of
copying the same policy into every system prompt.

## Context [#context]

Context is reusable reference material and memory.

Use it when an agent needs **background knowledge** repeatedly, such as:

* architecture notes
* domain terminology
* system boundaries
* durable preferences
* timeline events worth remembering

The current Context surface is memory-oriented, not just a static text list.
It supports four lanes:

* **Session** for the current working set
* **Semantic** for durable truths and preferences
* **Episodic** for timeline events and milestones
* **Procedural** for reusable workflows and runbooks

You can add entries, search them, delete them, and run **Compact** or
**Checkpoint** actions. Context can also be scoped to the current repo or
workspace session when that information is available.

## How to split responsibility [#how-to-split-responsibility]

Use this decision rule:

* if it teaches a repeatable workflow or specialty: **skill**
* if it constrains behavior or policy: **rule**
* if it explains the system or preserves useful memory: **context**

When those are separated cleanly, agent prompts stay smaller, more reusable,
and easier to maintain.

## Why this matters in practice [#why-this-matters-in-practice]

Without this split, teams tend to create one giant agent prompt that mixes:

* capabilities
* policies
* architecture notes
* temporary instructions
* memory that should have been stored elsewhere

That works briefly and then becomes difficult to reason about or update.

Mogplex works better when the roster entry stays readable and these supporting
libraries hold the reusable detail.

## A good composition pattern [#a-good-composition-pattern]

For a strong reusable agent:

1. define the core identity in the roster
2. attach shared rules for non-negotiable behavior
3. add skills for specialist workflows
4. use context for background knowledge or memory the task actually needs

That gives you a stable reusable agent without turning every task into prompt
surgery.