Core concepts

PromptBranch keeps prompt work organized without making you choose between quick edits and a reliable history.

Prompts, drafts, and versions

A prompt is the item in your library: it has a title, optional description, content, and organization such as tags or collections. Creating a prompt also creates its first version on the main variation.

Edit the displayed saved version as a draft. With autosave enabled, your in-progress work is retained against that exact version when you switch prompts or close the app. Press ⌘S on macOS or Ctrl+S on Windows and Linux, or choose Save changes, to update the displayed version without changing its ID, label, number, or place in the version tree.

Choose Save as new version when you want to preserve the displayed version and create a descendant. Include a short change note so you can understand the reason for that new version later.

Version numbers are display labels scoped to a variation. Deleting an older version leaves every surviving number unchanged, so gaps such as v1, v3, and v4 are normal. The CLI and MCP server return a stable versionId; use it when automation must keep referring to the same version record. Its content can be amended in the desktop app, so save a new version when content itself must stay as a historical snapshot.

Current version and variations

Each prompt has one current version: its preferred/default saved revision. This is what search, Quick access, the CLI, and the MCP server open or return unless a caller asks for a specific version or variation. Save changes never moves this pointer. Save as new version advances it when the displayed base is current; saving a descendant from another historical version preserves the current pointer until you deliberately change it.

Use Duplicate as variation to explore a different approach. A variation copies the version currently displayed, opens that new working version, and has its own numbered history. The current version remains unchanged until you deliberately set another version as current.

Runs, ratings, and notes

Every desktop model run records the model, output, timing, token usage, and estimated cost when available. You can compare runs, add a manual rating, or ask a connected model to judge an output. Notes capture context and findings for a prompt or a particular version.

Agent suggestions

Agents can read prompts, log runs and notes, and propose a variation. Their suggestion appears in the desktop app's Suggestions view as pending. It does not show up in ordinary searches or become current until you approve it. Rejecting a suggestion keeps it out of use while preserving the review record.

Tags, collections, and Trash

Use tags for labels you can reuse across prompts, and collections for groups of related prompts. Star important prompts for quick access. Deleting a prompt moves it to Trash, where you can restore it or permanently remove it when you are sure.

Next: manage versions and variations or connect an AI provider.