Prompt Manager — ComfyUI Custom Node

A node that manages a fully editable, categorized library of reusable prompt snippets, with a native multiline "Prompt" text box, presets stored on disk, per-section enable/disable, and queue-time randomization.

Installation

1. Copy the whole prompt_manager_node folder into ComfyUI/custom_nodes/.
2. Restart ComfyUI. Only dependency is aiohttp, which ComfyUI already ships with — nothing extra to pip install.
3. The node appears under category utils/prompt → 🗂️ Prompt Manager.

Structure:

prompt_manager_node/
├── __init__.py                 # node registration + WEB_DIRECTORY
├── prompt_manager_node.py      # Python node logic
├── server_routes.py            # /prompt_manager/presets/* API routes
├── presets/
│   ├── default.json            # shipped starter library
│   └── .last_used              # created automatically, tracks the active preset
└── web/
    └── prompt_manager.js       # custom UI injected into the node

Node widgets (top to bottom)

1. seed (native int widget) — seeds the queue-time randomization (see below).
2. Prompt (native multiline text box) — free text, always available. It is tied to a special, locked "Prompt" section: you can drag its tab to reposition it among the other sections (to control where its text lands in the final output), but it can't be renamed or deleted, and its content is not saved into the preset JSON (only its position is). It has its own accent color in the tab strip.
3. The custom Prompt Manager panel (tabs, toolbars, list/grid, preview).

UI features

Sections (tabs)

• Drag any tab (including the locked "Prompt" one) to reorder sections — this is the order used when building the final output.
• Click "+" at the end of the tab strip to add a new section.
• Each tab has a small 👁/🚫 icon: click it to enable/disable that whole section without touching its prompt selections. A disabled section is fully skipped from the output, whatever is checked inside it.
• The section toolbar's ✎ / 🗑 rename/delete the active section (disabled for the locked "Prompt" section; you must keep at least one regular section).

Selection model (unified)

• Clicking anywhere on a card/tile (list or grid mode) toggles whether that prompt is selected — this single flag is both "included in the final output" and "targeted by the bulk-action buttons" (delete/copy/cut/move). There's no separate checkbox anymore; the whole card is the clickable zone.
• The ◎ icon on a card is "select only this one": it selects that prompt and clears every other selection in the same section — handy for quickly swapping a single choice without hunting down what else was checked.

Two view modes (toggle with the ▦/☰ icon in the preset row)

• List — one row per prompt, shows the prompt text preview, good for reading/reviewing.
• Grid — compact tiles that wrap and re-flow automatically as you resize the node, bigger thumbnails, name only (no prompt-text preview) — better for browsing many reference images at a glance.

Both modes support drag-and-drop reordering (list order = output order) and the same click-to-select / ◎-to-solo-select / icon actions (✎ edit, 🖼 image, 🗑 delete).

Bulk actions (section toolbar, act on whatever is currently selected in the active section)

• ⌫ clear selection (this section only).
• 🗑 delete selected, 📋 copy, ✂ cut, 📥 paste (into the active section), ➡ move… dropdown to move selected prompts to another section.

Import / export

• ⇩ / ⇧ in the section toolbar export/import a single section as a small JSON file. Importing a section whose name collides with an existing one automatically appends (1), (2), etc., so nothing gets overwritten.
• The preset row (💾/✎/🗑 + dropdown) manages whole-library presets, saved server-side in the node's own presets/ folder (see below) rather than as browser downloads.

Presets folder

• 💾 asks for a name and saves the entire current library (all sections, all prompts, images included — but not the live "Prompt" text box content) as presets/<name>.json.
• The dropdown lists every preset in that folder; selecting one loads it immediately and remembers it as "last used".
• ✎ renames the selected preset's file, 🗑 deletes it.
• On ComfyUI startup, brand-new Prompt Manager nodes default to whatever preset was last used (via presets/.last_used), falling back to presets/default.json if none was ever selected. Note: ComfyUI typically caches node defaults per browser session, so if you switch presets mid session, already-open "add node" menus might not reflect it until you refresh the page — nodes you explicitly load from a saved workflow are unaffected either way, since they keep their own saved data.

Randomization at queue time

• The 🎲 icon in the section toolbar is a toggle, not an instant pick: when enabled (highlighted) for a section, that section gets exactly one random prompt selected automatically the moment you hit Queue Prompt — not while you're editing.
• The seed widget makes this reproducible: the same seed always reproduces the same picks for a given section (each section gets an independent but deterministic pick derived from the seed).

Live preview

• The bottom strip recomputes the exact output string as you edit — including the live "Prompt" text box content and whichever picks a random section currently holds.

Output format

Single STRING output (prompt), built by walking sections in their current (user-defined, draggable) order:

• An enabled locked "Prompt" section contributes the text box's content as-is.
• An enabled regular section contributes its checked prompts joined by ", ".
• Disabled sections are skipped entirely.
• Non-empty section contributions are joined by ". ".

Example: text box contains masterpiece, best quality; in Style you checked photorealistic, ultra detailed, 8k; in Camera you checked close-up shot; all three sections enabled, in that order →

masterpiece, best quality. photorealistic, ultra detailed, 8k. close-up shot

Data format (preset JSON)

{
  "viewMode": "list",
  "sections": [
    { "key": "__raw_prompt__", "label": "Prompt", "enabled": true, "locked": true, "randomizeOnQueue": false },
    { "key": "style", "label": "Style", "enabled": true, "locked": false, "randomizeOnQueue": false }
  ],
  "categories": {
    "__raw_prompt__": [],
    "style": [
      { "id": "p_ab12cd", "name": "Photorealistic", "prompt": "photorealistic, ultra detailed, 8k", "image": "data:image/jpeg;base64,...", "selected": true }
    ]
  }
}

This JSON lives in the node's hidden prompt_data widget, so a workflow save/reload keeps everything (it doesn't depend on the presets folder at all) — the preset picker is purely a convenience for reusing a library across different workflows or node instances.

Server API (used internally by the UI)

• GET /prompt_manager/presets → { names: [...], last: "name"|null }
• GET /prompt_manager/presets/{name} → preset JSON content
• POST /prompt_manager/presets/{name} (body = preset JSON) → save/overwrite
• POST /prompt_manager/presets/{name}/rename (body {new_name}) → rename
• DELETE /prompt_manager/presets/{name} → delete
• POST /prompt_manager/last_used (body {name}) → set the "last used" marker

Notes / known limitations

• The UI relies on node.addDOMWidget(...) and a computeSize-driven height to track the node's vertical size — both are ComfyUI frontend internals that can shift between versions. If vertical resizing stops filling the panel after a ComfyUI update, check that version's app.js/widgets.js for changes to this API.
• Queue-time randomization is implemented by wrapping app.queuePrompt. This is a common but unofficial pattern; if a future ComfyUI version renames or restructures that function, the wrap may silently no-op (the rest of the node keeps working normally, just without auto-randomization).
• Section add/rename/delete/preset-name prompts use native browser prompt()/confirm() dialogs for simplicity rather than a custom modal.
• Images are stored as base64 directly in the JSON (auto-resized to 220px client-side); a library with many images will make the preset file (and the saved workflow) proportionally larger.
• No external Python dependency beyond json/os/re (stdlib) and aiohttp, which ComfyUI already bundles for its own server.