linzhe001/data-prep

Data Prep

References

Read these first:

• ../../../.agents/references/workflow-guide.md
• ../../../.agents/references/language-policy.md
• ./references/dataset-stats.md
• ../../../PROJECT_STATE.json

When To Use

Use this skill for WF4 when the user wants dataset analysis, subset design, and data-pipeline preparation.

Required Work

1. Resolve dataset name, dataset path, and optional subset strategy from PROJECT_STATE.json and the user request.
2. Auto-detect dataset format and infer task type.
3. Produce the canonical stats for the detected task family.
4. Generate a reproducible subset strategy:
   • NVS or 3DGS: resolution scaling, scene selection, or point-cloud downsampling
   • detection: stratified subset indices
5. Write docs/Dataset_Stats.md using the canonical template.
6. Write the expected config artifact, such as configs/subset_config.json or configs/subset_indices.json.
7. Create or update the data pipeline script path expected by the canonical prompt.
8. Update PROJECT_STATE.json, especially dataset_paths, when appropriate.
9. Refresh CLAUDE.md so ### Dataset Paths reflects the resolved dataset addresses immediately after WF4.

Output Rules

• Use ./references/dataset-stats.md.
• Keep the context_summary, dataset format summary, full stats, subset strategy, and expected speedup.
• Dataset path synchronization into CLAUDE.md is required WF4 output, not an optional downstream refresh.
• Treat template wording as structure-only; localize headings and narrative text according to ../../../.agents/references/language-policy.md unless a field is explicitly English-only.

Codex Adaptation

• Treat natural-language requests as the canonical $data-prep flow.
• Preserve the original task-aware dataset logic, especially the NVS or 3DGS rule against random view dropping.
• Keep the original outputs and state-update behavior.

Execution Rule

Follow the local prompt, template, and language policy closely, especially the data-format detection and subset-strategy rules.