evaluating-instruction-following

What problem does it solve? After an image generation or editing step completes, there is no automatic way to know whether the result actually matches what the user asked for. This Skill provides a deterministic pass/fail gate that scores instruction adherence before a step is marked done, feeding failures back into the repair loop. ## Core Features & Use Cases - Heuristic Scoring: Combines goal-token matching, instruction-to-prompt token overlap, and denoise-range sanity checks into a 0-1 score with a 0.5 pass threshold. - Structured JSON Contract: Reads {instruction, result_meta} on stdin and emits {pass, score, reasons[], evaluator} on stdout, so it can be swapped for a vision-LLM evaluator later without interface changes. - Smoke Evals Included: Ships an eval runner covering pass, fail, and malformed-input cases. - Use Case: After a ComfyUI generation step produces an image for the instruction "generate a photo of a red fox in snow", run this evaluator to confirm the goal, prompt, and denoise settings align with the request before accepting the result. ## Quick Start Evaluate the last generation result against my original instruction and tell me whether it passes.