08. The LLM Prompts Used

This chapter walks through the design process behind this tutorial — from a first experiment that didn't need a DSL, through consulting an LLM for guidance, to the project spec that produced Slangify::Tutorial.

First experiment: bill extraction

Starting point: a folder of utility bills.

Ask Claude: "extract the data from these bills"
Result: bills.csv  ✓

It worked — but there was no need for a DSL. A plain LLM prompt handled the extraction cleanly. The lesson: reach for a DSL only when structure and validation matter downstream.

This used approx. 10 bills, from the same utility provider. Possibly, the quality of data extraction would reduce with a more mixed bag of source material, more items or variations in LLM interpretation (eg different LLM models) which would suggest a return to the DSL strategy.

Consulting an LLM: where does DSL + LLM shine?

A conversation with ChatGPT surfaced the right use case:

I want an LLM to extract structured information from messy text — but without hand-writing JSON schemas or fragile parsing logic.

Example input:

Jane booked a table for 4 at 7:30pm tomorrow at Bistro Verde

Desired output:

{
  "name": "Jane",
  "party_size": 4,
  "time": "7:30 PM",
  "restaurant": "Bistro Verde",
  "date": "tomorrow"
}

The DSL approach wins here because the output must be typed and validated — not just text.

Project spec

Use mi6 to scaffold a new Raku project Slangify::Tutorial in directory Slangify-Tutorial.

Set out the specification via openspec.

Make a DSL for this purpose using Raku Grammar and Actions (see examples at https://slangify.org). Declare any needed classes in the Actions file.

Make sure Grammar and Actions do the whole job so they can be used in the Grammar Editor / Slangify Playground.

Use JSON::Fast, Actionable, and LLM::Functions. Prompt for other Raku modules if needed.

Make a command Slangify-Tutorial that takes the text as an argument and optionally a filename for the JSON output.

Spec tweaks

• Add docs and tests.

• Use the standard Raku / IntelliJ .gitignore.

• Any class instantiated inside an Actions method (e.g. via .action($/)) SHALL be declared with a simple unqualified name (e.g. class Booking) rather than a fully-qualified namespace name (e.g. class Slangify::Tutorial::Booking).

• Make name cover names with titles, first name, middle name, last name, and so on.

Result

View in Slangify Playground