Pachinko pachinko
Pachinko pachinko
Design Notes on Useful AI

reuse / note 08

I rebuild the same prompt every time.

One-off chats are fine until the third time you assemble the same goal, constraints, examples, and tone from scratch.

Same ingredients, every time.

A good result depends on the same ingredients: the goal, the examples, the constraints, the tone, the source material, the rules for deciding. Rebuilding that in a chat box every morning is a tax on everything you make.

The tool throws away the valuable part.

The conversation ends and the setup goes with it. The output survives - the cheap part - while the expensive part evaporates. Thirty years of software trained users to expect their work to be saved. Chat somehow shipped without that. Spreadsheets have always had the fix: name the computation and let new inputs run through it.

What to build.

Let users turn a good exchange into a saved, named object: prompt, context, examples, and settings together, so new material runs through it without re-explaining anything. Make these objects visible and editable rather than buried in a settings menu. Give them versions. A setup that works gets tuned for months.

design principle

Turn repeated setup into a reusable object users can inspect, tune, and rerun.

one implementation

Pachinko's version: Functions store prompt, context, examples, and settings together. Drop new notes on one, batch with Map or Reduce, pipe a Feed into it, or export it for someone else to use.