Strata 提示词助手 / Prompt

If StrataT8Text is the raw model, this is the one with opinions. You type a rough idea, pick a task, and it comes back with something you can wire straight into a text encoder. It's the node most people actually install this pack for.

The six tasks

task is a fixed list, and each one is a system instruction the pack wrote - you can read them all in the source:

• expand - grow the input into a clearer, more concrete prompt without changing the subject.
• translate - accurate English translation, technical terms preserved.
• rewrite - clarity and specificity, constraints kept.
• image - a full English image prompt: subject, action, composition, setting, lighting, style.
• video - an English video prompt with visible action, camera movement, temporal progression and continuity.
• positive_negative - returns JSON with a positive and a negative string.

Every preset ends with some version of "output only the prompt", which is the whole fight with a chat model in a graph. And because the engine returns reasoning on a separate channel, the pack isn't left regex-stripping "Sure! Here's your prompt:" out of your conditioning.

Why bother

The KB's framing is a good one: on 2026's LLM-encoded bases you're already feeding a language model an instruction, so having a second language model write that instruction is just translation between two things that speak the same language (the KB's llm-in-comfyui notes § 1). It doesn't make your images better on its own. It removes the blank page - and it's good at the two jobs people are worst at: turning a vague vibe into specific visual nouns, and turning your own broken English into prompt-shaped English.

Inputs

• connection - from StrataT8Connection.
• text - your rough idea. This is the user message.
• task - the preset.
• template - an editable system prompt. If it's not blank, it replaces the chosen preset entirely. So a custom template is not an add-on, it's an override, and whatever you write is what the model gets. Keep the "output only the prompt, no preamble" instruction in anything you write, or you'll be back to fighting formatting.
• The usual sampling set: max_tokens 1024, temperature 0.7, top_p 0.95, top_k 20, seed, reasoning_effort, refresh.

Outputs

positive goes into your positive CLIP Text Encode and works for every task. negative only gets filled in by positive_negative; for the other five it comes back as an empty string, which is exactly what you want feeding a negative encoder that you're not customising. usage_json is token accounting for tuning.

That positive_negative mode is the sleeper feature, by the way. It requests a strict JSON object with exactly two required string fields named positive and negative, and the pack validates the JSON before returning it - wrong shape, and you get an error instead of a half-parsed prompt silently landing in your encoder.

Install

Same drill as the rest of the pack. Manager → search Strata-T8 (strata-t8, publisher t8star), or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Strata-T8.git
python -m pip install -r Comfyui-Strata-T8/requirements.txt

Then restart, open the Strata-T8 sidebar, save a profile, and select it in Connection. jsonschema is the only package you're adding; images and video reuse ComfyUI's own Pillow/PyAV, and the engine keeps its own Python and CUDA stack. No weights ship with the nodes - the portable runtime plus Strata-data is a separate, large download (IQ3_S is ~83.6GB, plus ~5.2GB of MTP tensors).

Where people get burned

The negative output being empty is not a bug. People wire both outputs, see nothing on negative for the image task, and go hunting. Switch to positive_negative.

task:video writes video-shaped prompts. It talks about camera movement and temporal progression - great for LTX/Wan-style video models, actively unhelpful for a still-image checkpoint that will treat "slow dolly in" as literal scene content. Pick the task for the model you're feeding, not for where the shot came from.

A custom template silently wipes the preset. The most common report I'd expect here is "I set task to image and it started writing paragraphs" - check template is actually empty.

It's a conversation, so context matters. A long text input plus a chatty system prompt can burn a lot of the 32K default context; the engine's context is a profile setting, not a node widget.

And it's still sampling. Same refresh caveat as every node in this pack: identical inputs and identical refresh means ComfyUI serves the cached answer, no model call. Bump refresh when you actually want a new take.