taxue-dsh-artisan

In DeepSeek Harness, going from a reference image to a generated one used to mean chaining two unrelated plugins and letting the model manually shuttle structured fields between them — fragile, slow, and full of silent failures. taxue-dsh-artisan folds the entire pipeline into a single plugin with four tools you can invoke individually or chain in sequence, eliminating the need for external glue scripts.

Each tool has a sharp focus. reverse_prompt reads an image and outputs a ready-to-paste generation prompt guided by nine quantified visual anchors (aspect ratio, mood, style, surface layer, composition, light/color/material, subject, visible text, negative constraints), with the ratio computed from real pixel dimensions and locked into the first sentence; an optional LLM re-writer converts raw observations into generation-ready instructions. extract_palette runs a deterministic pixel-level color extraction (K-Means in Oklab space) on PNGs, returning primary, secondary, and accent hex values with proportions — the same image always yields the same palette. optimize_prompt audits a candidate prompt locally for camera conflicts, style contradictions, and IP risk, then suggests single-variable fixes. generate_image renders via Google Gemini, OpenAI, ByteDance Seedream, xAI Grok, or a ChatGPT subscription, with optional reference-image side-by-side preview and image-to-image control; every output is saved to disk alongside a JSON sidecar containing the full prompt and metadata.

Built for solo developers, art directors, and prompt engineers working inside DeepSeek Harness who need to go from a client screenshot or mood board to a reproducible prompt, verify exact color values, and iterate across multiple image providers — all without juggling three plugins and hand-copying fields between them.