Parametric learning in your code harness.
machine docs tinker coding-agents.md
2.3 kB · 72 lines
Markdown

Using pi-tinker with coding agents #

Pi is the main interface. Other coding agents can run the same /tinker workflows through a small shell adapter.

Pi #

pi install git:github.com/gvkhosla/pi-tinker
pi

Then use /tinker commands directly:

/tinker demo
/tinker doctor
/tinker improve data.csv --goal "better answers" --budget demo

Pi also lets you select Inkling or a trained checkpoint with /model.

Claude Code, Codex, Cursor, Copilot, or Gemini CLI #

From this repository, replace /tinker with node scripts/agent-cli.mjs:

In Pi In another coding agent's shell
/tinker demo node scripts/agent-cli.mjs demo
/tinker doctor node scripts/agent-cli.mjs doctor
/tinker validate data/train.jsonl node scripts/agent-cli.mjs validate data/train.jsonl
/tinker improve data.csv --budget demo node scripts/agent-cli.mjs improve data.csv --budget demo

For example:

node scripts/agent-cli.mjs doctor
node scripts/agent-cli.mjs improve data.csv --goal "better support answers" --budget demo

The adapter prints the same report that Pi displays. Generated Python and JSONL are identical.

Prompt to give your agent #

Read AGENTS.md. Use pi-tinker to prepare and validate my data for Inkling.
Show me the generated files. Do not call the API or start training until I approve.
Keep the baseline and checkpoint eval at the same reasoning effort.

Safe workflow #

  1. Run doctor.
  2. Prepare and inspect the data.
  3. Create and inspect held-out eval examples.
  4. Validate the renderer and token masks.
  5. Ask before any command that uses the API.
  6. Run a two-step smoke test before a larger run.
  7. Compare the base model and checkpoint at the same Inkling effort.

The demo budget makes no API calls. The smoke, small, and real budgets do.

Agent instruction files #

The repository includes instructions for common agents:

  • AGENTS.md — canonical guide for all agents
  • CLAUDE.md — Claude Code
  • GEMINI.md — Gemini CLI
  • .github/copilot-instructions.md — GitHub Copilot
  • .cursor/rules/pi-tinker.mdc — Cursor

All agents should use official Tinker and Tinker Cookbook APIs and keep the generated code visible and editable.