Key concepts

A quick reference for the main concepts you'll encounter in Contextium. These building blocks work together to organise your knowledge and ground your AI agents.

Workspace

A workspace is the top-level container for everything your team works on. All your files, libraries, and team members live inside a workspace. Most teams have one workspace; larger organisations may have several.

Context library

A context library groups related files together — for example, one for customer-facing help articles and another for internal engineering docs. Every file belongs to one context library.

File

A file is a single document. You write in Markdown, and changes save automatically. Every file has a version history. Files are the primary unit of context — when an AI agent is grounded, it reads your files. Files are stored within a Context library, and provide detailed information about a topic.

Skill

A skill is a reusable instruction set stored in your workspace. Skills tell an AI agent how to do something — for example, "follow our brand voice" or "use our API error format". You write a skill once and any agent or team member can load it. A generic skill can be combined with a file to build more specific context.

Agent

An agent definition in Contextium describes a persona or role for an AI assistant — its name, purpose, and which skills it should use. Defining agents in Contextium means every team member gets the same agent behaviour, grounded in the same context.

Tag

Tags are flat, untyped labels on files. You add a tag by typing #value directly inside a file's content — for example #draft or #engineering — and Contextium reconciles the file's tags from its body when you save. A file can carry as many tags as you like, so the same file can be found from several angles at once. There are no tag types, categories, or hierarchies.

Identity

An identity captures the voice, tone, and perspective that AI should write from. It stores details like role and context, background, intended audience, voice descriptors, writing samples, and rules for what to always or never do. One identity can be attached to a workflow so every agent in that workflow writes consistently from the same perspective — whether that's a personal brand, a company tone, or a founder who speaks for both.

Identities come in three types:

• Person — an individual voice (e.g. a founder, consultant, or team member).
• Company — a brand or organisation tone.
• Hybrid — an individual who also represents a company (e.g. a founder writing as themselves but for a business audience).

Identities can be user-scoped (personal, only visible to you) or workspace-scoped (shared across the team).

Ruleset

A ruleset is a named collection of rules — short instructions your agents follow. You write a rule once, attach the ruleset to as many workflows as you like, and every agent working on them follows it. Attachment is a live link, not a copy, so editing a ruleset changes every workflow using it rather than leaving stale duplicates behind.

Every rule has one of four types:

• Must — comply; if the agent cannot, it stops and asks rather than working around it.
• Must not — never do this.
• Prefer — use this, but fall back to the next best option if it is genuinely unavailable.
• Consult — read a named context library when a stated condition applies, and not before.

Consult is the one worth knowing about. Rather than pasting a whole design system into your rules, keep it in a library and add a consult rule pointing at it. The pointer is always present and costs almost nothing; the library is only fetched when the work calls for it.

Each workspace has one baseline ruleset that applies to every workflow automatically and cannot be detached. Keep it small — it is for safety and compliance ("never commit credentials"), not conventions. A baseline prohibition can never be overridden by a workflow; a baseline requirement or preference can.

Rules also carry an enforcement level. Advisory means the rule is included in every agent's instructions — a strong instruction, but not a guarantee. Checked means something actually blocks the behaviour, such as a linter rule or a CI check, and can only be set when a real mechanism is linked.

Workflow

A workflow is a saved collection of resources defined by tags. When you load a workflow, Contextium finds all files, skills, and agents that match the workflow's tags and makes them available to your AI assistant in one click. Workflows are the fastest way to ground an agent in a specific task — instead of loading resources one by one, you load a whole context in a single action.

Roles

Every workspace member has a role that controls what they can do:

• Owner — Full control, including billing and workspace deletion. One per workspace.
• Admin — Can manage members, settings, and all content. Cannot access billing.
• Editor — Can create, edit, and delete content. Cannot manage members.
• Viewer — Read-only access.

For the full permissions matrix, see Role permissions.