Skip to content

OVERVIEW.md

Latest commit

 

History

History
91 lines (68 loc) · 3.35 KB

OVERVIEW.md

File metadata and controls

91 lines (68 loc) · 3.35 KB
Raw

Overview

Toolkit for building prompt-tuning datasets from your own Cursor history — not a bundled corpus. You harvest locally; scripts and docs are what belong in git.

What it does

  • Collect agent JSONL from ~/.cursor/projects/*/agent-transcripts (parent sessions and subagents)
  • Catalog sessions in a local manifest (paths, repo hints, tags)
  • Normalize to flat turns.jsonl with tool/skill metadata
  • Split by session into eval / pool / discard (pnpm split)
  • Aggregate patterns (pnpm insights) and draft rules/skills (pnpm suggest-artifacts)
  • Install distilled artifacts into target repos (pnpm install-artifacts → <target>/.agents/)

Bundles are your repo_hint values (from workspace slugs), not a fixed list. After split: pnpm suggest-artifacts -- --list or pnpm insights.

Repo layout

agent-prompt-tuning-lab/
  scripts/
    harvest.sh              # host / devcontainer archives + unpack
    normalize.mjs           # --source host|devcontainer|manual|all
    split.mjs               # eval / pool / discard by session
    seed-manifest.mjs
    tag-manifest.mjs
    insights.mjs
    suggest-artifacts.mjs   # Phase 4 LLM drafting
    install-artifacts.mjs
    lib/                    # artifact context + LLM clients
  .cursor/skills/           # harvest workflow skill
  .cursor/rules/            # privacy + normalize conventions
  docs/                     # pipeline, workflow, artifacts, schema
  data/                     # data/** ignored except .gitignore allowlist

Raw unpack layout:

data/raw/host/<YYYYMMDD>/<workspace-slug>/<session-id>/…jsonl
data/raw/host/…/<session-id>/subagents/<subagent-id>.jsonl
data/manifest.jsonl         # local catalog

Harvest

Full extract (pnpm harvest:all):

  1. Host — all ~/.cursor/projects/*/agent-transcripts (--all), including subagents/*.jsonl.
  2. Devcontainer — optional second pass; filters workspaces by slug or CURSOR_REPO_NAMES.

Selective harvest:

pnpm harvest:host -- --all --unpack
CURSOR_REPO_NAMES=my-app,other-repo pnpm harvest:host -- --unpack
pnpm harvest:devcontainer -- --unpack
Variable Default Purpose
CURSOR_PROJECTS_HOST $HOME/.cursor/projects Host projects root
CURSOR_REPO_NAMES (none) Comma-separated repo name fragments

Details: PIPELINE.md.

Gold sessions

Tag exemplars locally, then record ids in GOLD_SESSIONS.md:

pnpm tag-manifest -- --tag gold --session-id <uuid>
pnpm tag-manifest -- --tag gold --repo <repo-hint> --limit 5

Roadmap

Phase Status Command
1 Harvest + manifest Done pnpm harvest:all, pnpm seed-manifest
2 Normalize turns Done pnpm normalize
3 Split Done pnpm split — eval / pool / discard by session, tags, repo
4 Suggest artifacts Done pnpm suggest-artifacts — Grok, Cursor SDK, or IDE prompt

Phase 3 does not split by Composer vs auto model (not in Cursor exports). Splits are by session, tags, and repo_hint.

Related docs