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 (v0.3.8)​

The official AILANG teaching prompt is maintained at:

prompts/v0.3.8.md​

This prompt is:

  • Validated through eval benchmarks - Tested across GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5
  • Up-to-date with v0.3.8 features - Record updates, auto-import prelude, anonymous functions
  • Versioned with SHA-256 hashing - Reproducible eval results
  • Actively maintained - Updated as language evolves

Quick Reference​

Current version: v0.3.8 (AI Usability Improvements)

What works in v0.3.8:

  • 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 - NEW in v0.3.8!
  • Record update syntax {base | field: value}
  • Auto-import std/prelude - No imports needed for comparisons
  • Anonymous functions func(x: int) -> int { x * 2 }
  • Numeric conversions intToFloat, floatToInt
  • Effects: IO, FS, Clock, Net
  • Type classes, ADTs, pattern matching
  • REPL with full type checking

What doesn't work yet:

  • Pattern guards (parsed but not evaluated)
  • Error propagation operator ?
  • Deep let nesting (4+ levels)
  • Typed quasiquotes
  • CSP concurrency

For complete details, see prompts/v0.3.8.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/v0.3.8.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/v0.3.8.md]

Now write the code.

For Eval Benchmarks​

The eval harness automatically loads the correct prompt version:

# benchmarks/example.yml
id: example_task
languages: ["ailang", "python"]
prompt_files:
  ailang: "prompts/v0.3.8.md"
  python: "prompts/python.md"
task_prompt: |
  Write a program that [task description]

See benchmarks/README.md for details.

Current Prompt​

Version: v0.3.8 - View full prompt

Features:

  • Multi-line ADTs: type Tree = | Leaf | Node (v0.3.8)
  • Record updates: {base | field: value} (v0.3.6)
  • Auto-import std/prelude (v0.3.6)
  • Anonymous functions: func(x: int) -> int { x * 2 } (v0.3.5)
  • Numeric conversions: intToFloat, floatToInt (v0.3.5)
  • Full module system with effects

Why prompt quality matters:

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

Eval Results​

Current success rates (v0.3.8 prompt on v0.3.8):

  • Overall: 49.1% AILANG success rate (vs 82.5% Python baseline)
  • Claude Sonnet 4.5: 68.4% best performer
  • Gemini 2.5 Pro: 65.8%
  • GPT-5: 63.2%

Improvement trajectory:

  • v0.3.7: 38.6% → v0.3.8: 49.1% (+10.5% improvement)
  • Fixed benchmarks: pattern_matching_complex, adt_option, error_handling, and more

See Benchmark Dashboard for detailed metrics.

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. Update the prompt at prompts/v0.3.8.md (or create new version)

  4. Document changes in prompts/versions.json (future enhancement)

See Also​

Last updated: October 15, 2025 for v0.3.8/v0.3.8