Prompt Quill Sailing to Conditioning

If you've seen Prompt Quill Sailing and thought "yes, I want to drift through a 3.2M-prompt vector store, but I don't want to drag two extra CLIP Text Encode nodes around" - this is your node. Prompt Quill Sailing to Conditioning is the exact same exploration feature, with one addition: it takes a clip input and bakes the sailed prompt and its negative straight into CONDITIONING tensors. Prompt strings out one side, ready-to-sampler conditioning out the other.

The name is a mouthful, but the value is real. Sailing is the pack's standout feature: instead of rewriting your prompt, the server keeps a "journey" on its side and each call fetches a prompt as far from the last one as possible, so consecutive generations wander off into prompt-space you'd never explore by hand. This conditioning variant just makes that loop tighter - no intermediate text-encoding steps, fewer nodes between you and the sampler.

How it works

The node POSTs your seed prompt plus all the sailing flags to the server's /get_next_prompt endpoint, gets back a new prompt and negative, then encodes both with your clip via clip.tokenize / clip.encode_from_tokens. Like the plain Sail node, it returns NaN from IS_CHANGED, so it never caches - every queue run is a fresh step in the journey. The raw strings come out alongside the conditioning, which is handy when the sail gets weird and you want to see what you actually prompted.

Inputs and outputs

The ones that matter:

• prompt - your voyage's starting point.
• reset_journey - the classic trap. Set true for exactly one image, then back to false, to restart the journey server-side. ComfyUI is stateless, so this toggle is the only "new voyage" button.
• distance - jump size per sail (1–10000, default 20). Higher drifts further and weirder.
• use_all_trip - pins the journey close to your seed prompt, so the "trip" stays themed instead of wandering.
• clip - from your checkpoint loader; the encoding is done with your model's own text encoder.

Everything else - add_style/style, add_search/search, add_negative/negative, rephrase/rephrase_prompt, summary, unload_llm - is an optional toggle-plus-text that you only touch when you want to bias the sail.

Outputs: Prompt and NegativePrompt (strings), plus Prompt Conditioning and Negative Conditioning for the sampler's positive and negative inputs.

Install

ComfyUI Manager → "Install via Git URL" → https://github.com/osi1880vr/prompt_quill_comfyui, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/osi1880vr/prompt_quill_comfyui

No pip installs, no model downloads from the pack - it's a one-file client using only requests. The real dependency is the Prompt Quill server from osi1880vr/prompt_quill, which does the LLM + vector-store work and wants a GPU with headroom (people report ~16GB to run it alongside ComfyUI on one box).

Troubleshooting

• It keeps sailing the old route. You skipped the reset_journey true-then-false cycle. Easiest gotcha in the pack to trip over.
• "Something went wrong, did you install Prompt Quill?" - the node's literal error string for a failed HTTP call. Server down, wrong port, unreachable host.
• Conditioning is a snapshot of the journey. Re-queuing the same workflow gives a new prompt every time (no caching, by design). If you need the exact same conditioning again, save the raw Prompt string and feed it through a normal CLIP Text Encode instead.
• Know your encoder. The encoding uses your checkpoint's CLIP, and Prompt Quill's store is largely SD1.5/SDXL-era data. On CLIP-encoded models (SDXL-lineage, Illustrious, Pony) it sings; on LLM-encoded 2026 models the verbose output still helps as an instruction, but the long negative may do little. When in doubt, glance at the raw string before you judge the node.

The project went quiet in 2025, so don't expect rapid fixes - but as shipped, this is the most workflow-friendly node in the pack.