Replicate local workflow

Use Replicate's official skills for model-specific work. This skill adds the local workflow around them: planning batches, preserving prediction records, downloading every output, naming files, embedding image metadata, reviewing candidates, and archiving results.

Before running

1. Confirm the requested output directory. If none was given, use outputs/<concise-task-slug>/ under the current project.
2. Confirm the model, prompt or inputs, output count, format, and any cost or quality constraint that is unclear.
3. Show the user the prompts and batch plan before creating predictions when the task is exploratory, expensive, or asks for multiple variants.
4. Check that REPLICATE_API_TOKEN exists without printing its value.
5. Load the relevant installed Replicate skills:
   • find-models to search and inspect schemas
   • compare-models to evaluate candidates
   • run-models to create and monitor predictions
   • prompt-images or prompt-videos for media prompting
6. If those skills are unavailable, install them globally with: