Prompt Deck

Every prompt feeder in ComfyUI has the same dirty secret: it's blind. You point it at a text file, set it to run line 531, and you have no idea what line 531 actually says until the image comes out wrong. Prompt Deck (PromptDeck from omervaner/ComfyUI-PromptDeck) fixes that: it renders the whole file inside the node, highlights the line that's running, marks what's next, and lets you jump around. If you batch-generate from a prompts file at all, this is the one you'll actually reach for.

What it is and where it fits

It's an output node that reads a .txt file with one prompt per line, skips blank lines and # comments, and feeds the current line's text downstream. You wire its text output into the text input of a CLIP Text Encode node (or Flux's conditioning encoder - it just spits out a string). It's built for the "run a deck of prompts" workflow: character variants, style tests, or a few hundred lines of A/B prompts. Set line_index's control to increment, queue 20 runs, and you get 20 different prompts - the same control_after_generate machinery (fixed / increment / decrement / randomize) you already know from the KSampler seed widget.

How it works

The Python side is dead simple: load the file, filter out disabled and bad-rated lines, then take line_index modulo the count of runnable lines - any value is safe and it wraps, so randomize gives you random lines for free. Then it applies text surgery in a fixed order: strips leading list numbering (12.), cuts the head up to your cut_to markers, cuts the tail from your cut_from markers, does a search/replace, and finally wraps the result in pre_text + line + after_text. The node's IS_CHANGED keys off the file's mtime and size (plus its ratings sidecar), so editing the txt mid-session is fine - the next run picks up the change.

The list renders inside the node with numbered rows, a highlighted running line, and a header showing position / total · word count; it even resizes with its rows. Hover a truncated line to see the full prompt in a floating box. Click any line to jump to it, hold a button to pin and repeat a line until you release it. Ratings (★ good / ✕ bad) write to a sidecar yourfile.txt.deck.json next to your prompts, keyed by line text rather than line number, so they survive restarts and file reordering.

The inputs that matter

• file_path - the path to your .txt, one prompt per line. The only thing you strictly have to set.
• line_index - position among runnable lines, with control_after_generate for increment / decrement / randomize / fixed. Also a saved widget, so your position survives a restart.
• pre_text / after_text - text wrapped around each line. Handy for a shared quality prefix or a fixed style suffix without editing the file.
• cut_to / cut_from - comma-separated markers that trim a line's head or tail. cut_from: cinematic, film grain drops style tails from every line in one go.
• search / replace - plain text substitution.
• rating_filter - all, good only, or hide bad, once you've starred your favorites.

The two outputs are text (the composed prompt - wire it into your text encoder) and filename_slug (a filesystem-safe slug, handy for a Save node's filename_prefix so images are named after their prompt). A detail worth calling out: the actual composed prompt gets written back into the read-only resolved box after each run, so the real text lands in your workflow JSON and PNG metadata instead of the stale greyed-out string in a connected prompt box.

Installing it

Zero dependencies - no pip, no requirements.txt, just pure Python stdlib plus one JS file. Either use ComfyUI Manager (search "ComfyUI-PromptDeck") or:

cd ComfyUI/custom_nodes
git clone https://github.com/omervaner/ComfyUI-PromptDeck

Restart ComfyUI, then add it via Add Node → PromptDeck → Prompt Deck (search "prompt deck"). The node runs standalone as an output node, so you can test the feeder without loading a single model.

Gotchas

• "File not found" means the path is wrong - backslashes are fine on Windows, just point at the real file.
• line_index counts runnable lines (enabled + passing the rating filter), not raw file lines - so the widget number won't match your editor's once you've disabled or rated lines out.
• resolved and disabled_lines are JS-managed - don't edit them by hand.

It's a niche tool with a genuinely small footprint (don't expect a big community behind it), but the README is thorough and the code is tidy. If you only ever type one prompt at a time, skip it. If you run hundreds from a file, this is the feeder that stops making you guess.