Project design packs

mlx-agent blueprint turns a short project goal into a guidance-only MLX project design pack (markdown + JSON). Use it so downstream agents can plan quantization, training loops, LoRA, and study materials before any scaffolding or training.

It does not scaffold repositories, download weights, or run training.

Quick start

python3 scripts/mlx-agent blueprint \
  --goal "On-device legal OCR assistant" \
  --modality document-vision \
  --memory-gb 64

python3 scripts/mlx-agent blueprint \
  --goal "Meeting ASR notes" \
  --modality audio \
  --notes "must stay offline"

# Render only (no files)
python3 scripts/mlx-agent blueprint --goal "Billing helper" --no-write

Packs land under ./mlx-blueprints/<goal-slug>-<timestamp>.md with a matching .json sidecar. Pass --project DIR to change the project root, or --json for machine output.

Brief fields

Flag	Purpose
`--goal`	Required one-line project goal
`--modality`	Repeatable foundation: `audio`, `video`, `document-vision`
`--memory-gb`	Optional host memory budget (shapes quant guidance)
`--notes`	Free-form constraints echoed into next steps

Pack sections

Goal and brief metadata
Recommended stack path (research → runtime preference → adopt → wire)
Quantization ideas (modality- and memory-aware)
Training loop sketch (dry-run / LoRA-first; no execution)
LoRA / adapter notes
Experimental MTX notes (optional research, not production defaults)
Study materials
Next steps

Safety

Writes only inside <project>/mlx-blueprints and refuses symlink escape. No Hub fetches, downloads, or training. Distinct from the dataset blueprint section inside research packs (blueprint.py).