BigPlayer Split Prompt

BigPlayer Split Prompt is the pack's nod to a prompting style that most of the 2026 ecosystem has quietly moved past - and that's fine, because the models that still want it really want it. It reads the shared session from BigPlayer Natural Language Root and emits five strings: text_l_positive, text_g_positive, text_l_negative, text_g_negative, and comments.

The L/G convention is the SDXL-era split: text_l_* carries local subject and content detail, text_g_* carries broader mood, composition, and global styling. It's the same L/G idea behind the two-encoder conditioning paths people used for regional and compositional control back when SDXL shipped. In this pack's prompt fragments, the LLM is explicitly told to keep the local content in text_l_* and the global style in text_g_*, and not to duplicate the full positive into both channels unless the chosen checkpoint prefers that fallback.

How it works

Same machinery as every BigPlayer module: the root makes one provider call covering all attached modules, and Split Prompt resolves only its own validated slice (split_prompt) from the shared session. The provider must return exactly those five fields or the whole result is rejected locally. Wire the L and G channels into whatever conditioning path your model uses - typically the separate L and G text encoders in an SDXL-style pipeline.

The honest take

Here's where I'd push back a little. A lot of current models encode a single prompt - there's no L/G conditioning path to split into, and the whole two-channel dance is dead weight on them. If you're running a modern single-encoder checkpoint, BigPlayer Basic Prompt does the same job with less to wire. Reach for Split Prompt when you specifically run an SDXL-lineage model or a workflow with separate L/G encoder inputs. The node is a faithful implementation of a real convention; it's just a convention that isn't the default anymore. If the comments field says the LLM chose to duplicate the prompt into both channels "as a fallback," it's because the checkpoint it was told about prefers it - that's the model talking, not the node being lazy.

Install and troubleshooting

Same pack install: ComfyUI Manager, search "BigPlayer Prompting", or git clone https://github.com/josh-ent/comfyui-bigplayer-llm-nodes into ComfyUI/custom_nodes, then restart. Dependencies are just httpx and pydantic; no model downloads, but you need a Grok API key for the root to call anything.

Errors here are almost always upstream: the root wasn't configured with a working provider, or the session doesn't include the split_prompt capability because this module wasn't attached (directly) to the root's session output when the graph ran.