AuK Instruction Builder X

Why you'd reach for it

AuK's tasks are template-driven: "Say the following with the same voice: '{text}'", "Replace '{original}' with '{replacement}'", "Add '{text}' before '{anchor}'". The model was trained on those shapes, so a request that looks similar to you but isn't shaped like one of them will underperform - and you'll blame the model.

AuK Instruction Builder X removes the guesswork. You pick a task from a dropdown, the node reveals that task's fields, you fill them in, and the output is the exact canonical sentence. Wire it into the instruction input of AuK Instruction Encode X and you're working in the shapes the model actually knows.

How it works

The node has exactly one input, task, and it's a dynamic combo - which is the interesting part mechanically. Instead of a fixed form, picking a task reshapes the node: each task in the catalogue declares its own slots (a text to speak, an original and replacement word pair, an anchor phrase, and so on), and those become visible widgets with that task's defaults already in them. That's why the node face looks different depending on what you selected, and why you don't see a dozen irrelevant fields at once.

The catalogue covers the whole documented task set - voice-description speech, voice cloning, whisper conversion, speech content editing (replace, insert before, insert after, remove), lyric editing, emotion and timbre, pitch, speed, volume, accent removal, nonverbal sounds, enhancement, and speaker/music separation. One task, one template, one output.

The output is a plain STRING, and you have two ways to use it:

1. Wire it into Instruction Encode's instruction input. This is the intended path, and it's the one where a connected string overrides the typed widget.
2. Read it and rewrite it. The generated text lands in the node's output, so you can copy the shape and then edit the wording - useful because the model accepts free text as long as it matches a known instruction pattern.

Typical wiring:

AuK Instruction Builder X (STRING) -> AuK Instruction Encode X (instruction)

When to use the builder and when not to

If you're doing a straightforward replace, insert or clone, use the builder - it's faster than remembering punctuation and quoting rules, and quoting matters (the anchors have to actually occur in the recording). If you're experimenting with something the catalogue doesn't cover, write the sentence yourself and match the template style closely.

There's a third option worth knowing about: AuK Prompt Enhance X does the same classifying job but from loose prose, using a local language model. So you have a manual builder for precision and an automatic enhancer for speed, and the builder is the better one when you know exactly what you want.

Install

ComfyUI Manager → search WorkflowX Configurator, or:

cd ComfyUI/custom_nodes
git clone https://github.com/haroonaslam/WorkflowX-Configurator
cd WorkflowX-Configurator
pip install "transformers>=5.3,<6" safetensors PyYAML tqdm

Given it's just string assembly, this node itself needs nothing - no model, no weights, no GPU. It's the ones downstream that need the AuK checkpoints. But it only appears in the menu once the pack loads, so the usual applies: install, restart, hard-refresh the browser. The dynamically-shaped node face is a frontend feature, and a stale browser bundle can leave you with a task dropdown that doesn't reshape anything.

Where it goes wrong

A wired string beats your typing. Once the builder is connected, edits to Instruction Encode's text box do nothing. It's the most common "the node is broken" report in any suite with this pattern, and it isn't broken.

Task mismatch on a connected script. If you generate an instruction for the wrong task - cloning when you meant a replace - nothing errors; you just get speech that follows the instruction you actually gave. Read the builder's output once before queueing the first time.

No output connection, no effect. The builder's string only matters if it reaches Instruction Encode. A dangling output is silent.