Skip to main content

AI Prompt Guide: Teaching AILANG to Language Models

Purpose: This document points to the canonical AILANG teaching prompt for AI models.

KPI: One of AILANG's key success metrics is "teachability to AI" - how easily can an LLM learn to write correct AILANG code from a single prompt?

Canonical Prompt (v0.4.4)​

The official AILANG teaching prompt is maintained at:

prompts/v0.4.4.md​

This prompt is:

  • Validated through eval benchmarks - Tested across GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5
  • Up-to-date with latest features - Record updates, auto-import prelude, syntactic sugar
  • Versioned with SHA-256 hashing - Reproducible eval results
  • Actively maintained - Updated as language evolves

Quick Reference​

Current version: (Teaching prompt: v0.4.4)

Core Features:

  • Module execution with effects
  • Recursion (self-recursive and mutually-recursive)
  • Block expressions ({ stmt1; stmt2; result })
  • Records (literals + field access + updates)
  • Multi-line ADTs type Tree = | Leaf | Node
  • Record update syntax {base | field: value}
  • Auto-import std/prelude - No imports needed for comparisons
  • Syntactic sugar: :: cons, -> function types, f() zero-arg calls
  • Effects: IO, FS, Clock, Net, Env
  • Type classes, ADTs, pattern matching
  • REPL with full type checking

Known Limitations:

For complete syntax guide, see prompts/v0.4.4.md

Using the Prompt​

For AI Code Generation​

When asking an AI model (Claude, GPT, Gemini) to write AILANG code, provide the full prompt from prompts/v0.4.4.md.

Example usage:

I need you to write AILANG code to solve this problem: [problem description]

First, read this AILANG syntax guide:
[paste contents of prompts/v0.4.4.md]

Now write the code.

For Eval Benchmarks​

The eval harness automatically loads the correct prompt version from prompts/versions.json:

# benchmarks/example.yml
id: example_task
languages: ["ailang", "python"]
# Prompt version is auto-resolved from prompts/versions.json
task_prompt: |
  Write a program that [task description]

See benchmarks/README.md for details.

Current Prompt​

Version: v0.4.4 - View full prompt

Core Features Documented:

  • Multi-line ADTs: type Tree = | Leaf | Node
  • Record updates: {base | field: value}
  • Auto-import std/prelude
  • Syntactic sugar: :: cons, -> types, f() zero-arg calls
  • Full module system with effects (IO, FS, Clock, Net, Env)
  • Pattern matching, recursion, type classes

Why prompt quality matters:

  • Better AI code generation
  • Reproducible eval results
  • Consistent teaching across models

Eval Results​

Current success rates:

  • See Benchmark Dashboard for latest metrics
  • Best model: Claude Sonnet 4.5 (consistently highest success rates)
  • Results updated after each release

Key Insights:

  • Teaching prompt quality directly impacts AI success rates
  • Multi-model testing reveals universal vs model-specific patterns
  • Iterative prompt improvements correlate with better code generation

Contributing Improvements​

If you find ways to improve the AILANG teaching prompt:

  1. Test your changes with the eval harness:

    ailang eval --benchmark all --model gpt-4o-mini

  2. Measure impact:

    tools/compare_prompts.sh old_version new_version

  3. Create new prompt version following the versioning system in prompts/versions.json

  4. Document changes with SHA-256 hash and notes in prompts/versions.json

See Also​