Core concepts

Specify's features are built on a handful of concepts. This page explains what each one is and how they connect.

How it fits together#

1. Your team and documents come together in a workspace.
2. A document project ties a workspace folder to a GitHub source.
3. The agent follows a document skill to generate docs from the code.
4. When the source changes, proposed updates are prepared for the affected docs.
5. Your team checks progress and proposed updates in the Inbox, and coding agents read the same context through MCP.

Throughout this process, AI uses Workspace RAG to find your team's docs and files as evidence, and for work that needs a reference standard, such as regulatory documents, it cites the standards library clause by clause.

Workspace#

The unit in which your team shares docs, conversations, decisions, and code history. Members and roles, connected tools, and your plan all belong to a workspace. Data is isolated per workspace.

Node#

The core unit of content in a workspace. Documents, folders, and agent rules are all nodes, organized in a tree. A document project is connected to a specific folder node.

Document project#

The unit that analyzes a GitHub repository to create and maintain developer docs such as architecture docs, onboarding guides, and API specs. A document project has:

• Source: The GitHub repository and branch used as evidence for the docs
• Document skill: The writing procedure that defines which docs to write and the discipline to follow
• Automation settings: Whether to trigger on push, the default output folder, and the project style guide

In the project list, each project shows a current, writing, or needs attention status.

Skill#

The procedure and conditions of use an agent follows while working. In a document project, the document skill provides the writing discipline used to generate docs, and auxiliary skills provide reference procedures that help the document skill. You can select up to 5 auxiliary skills.

Skills fall into two types based on how they update docs.

Proposed update#

A suggestion Specify prepares when a code change means a document needs to change. Each proposed update includes the evidence for the change and a before-and-after view. New projects start without automatic updates, and how updates are applied follows your project settings and the applicable policy. For details, see Review and apply proposed updates.

Inbox#

Where the progress and results of background tasks collect. Even after you close the document generation screen, you can keep checking completed, failed, and try-again options in the Inbox. When an agent is waiting for your answer, you can also pick up from the Inbox and the thread.

Thread#

A unit of conversation with an agent. Questions and answers, and the work the agent performs, are recorded in the thread. If the agent asks a question during document generation, open the thread and answer it to continue generation.

Workspace RAG#

RAG (retrieval-augmented generation) is an approach where AI first searches for relevant materials and uses them as evidence before answering or writing a document. Specify indexes your workspace's documents, uploaded files, and messages from connected Slack public channels, and finds the parts it needs in chat and document writing. Search is isolated per workspace and runs only after checking member permissions. See Workspace RAG.

Standards library#

A knowledge source for searching official standards such as IEC 62304 and ISO 14971 clause by clause. For each section of a document, the agent searches the relevant clauses to check against, and leaves the clause number and edition as citation links. Standards search is rolled out in stages, subject to license verification. See Standards library.

Models and reasoning effort#

Chat and agent tasks run on one of the Claude or GPT family models. Each model differs in quality, speed, and cost, and raising the reasoning effort lets the model reason more deeply about complex tasks. See Models overview.

MCP#

The Model Context Protocol is an open standard for connecting AI agents to external tools and data. Specify provides an MCP server so agents such as Claude Code, Codex, and Cursor can search and read your team's docs within the scope your workspace permissions allow. See MCP overview.