Local toolkit to harvest Cursor agent transcripts, normalize turns, split for eval/pool, and distill rules and skills for other repos. No bundled dataset — you run the pipeline on your machine.
git clone <your-fork-url> agent-prompt-tuning-lab && cd agent-prompt-tuning-lab
chmod +x scripts/*.sh scripts/*.mjs
pnpm harvest:all && pnpm seed-manifest && pnpm normalize && pnpm split
pnpm insights # see repo_hint names in output
pnpm suggest-artifacts -- --list # bundles from your splits
pnpm suggest-artifacts -- --bundle <repo> --llm prompt # see docs/PROMPT_MODE.md
cp data/bundle-targets.example.json data/bundle-targets.json # map repos → local paths
pnpm install-artifacts -- --bundle <repo> --include-personal| Doc | Contents |
|---|---|
| Overview | Purpose, repo layout, roadmap |
| Pipeline | Harvest → normalize → split → suggest-artifacts |
| Workflow | Weekly cadence, porting lessons to other repos |
| Extraction | Rule vs skill, insights, artifact loop |
| Prompt mode | --llm prompt + ingest workflow |
| Bundle targets | Map repos to local project paths |
| Artifacts | Per-project rules and skills |
| Schema | Manifest and turn fields |
| Gold sessions | Tagged exemplar session ids |
| Contributing | PR guidelines |
Do not commit raw transcripts, processed turns, or data/manifest.jsonl.