AILANG Code Writing

Write correct AILANG code using the live MCP server as the canonical reference.

Quick Start

ailang run --caps IO --entry main solution.ail   # Run with effects
ailang repl                                       # Interactive development
ailang check solution.ail                         # Type-check only
ailang verify solution.ail                        # Prove contracts with Z3

When to Use This Skill

• User asks to write AILANG code, fix syntax errors, or run programs
• User asks about AILANG features, capabilities, or version
• User wants to add Z3 verification or contracts

Source of Truth: the AILANG MCP Server

mcp.ailang.sunholo.com is the canonical, live, version-locked source of truth for AILANG syntax, stdlib, examples, and limitations. Prefer it over any embedded reference in this file.

If the user's harness supports remote MCP (Claude Desktop, Cursor, Cline, Continue, Claude Code), have them add it once:

{
  "mcpServers": {
    "ailang-docs": {
      "url": "https://mcp.ailang.sunholo.com/mcp/",
      "transport": "streamable-http"
    }
  }
}

Or use the one-click button at ailang.sunholo.com.

Then the agent has typed tools to query AILANG knowledge directly:

Tool	When to call
prompt_get(for_version, kind="agent")	Need full syntax reference (replaces reading resources/syntax_quick_ref.md)
stdlib_modules(for_version)	Discover what's available in stdlib
stdlib_module(name, for_version)	Look up exports of a specific module before importing
stdlib_search(query, for_version)	Find a function by name/keyword
examples_for_concept(concept, for_version)	Find a working example of a feature
example_get(path, for_version)	Read full source of a specific example
limitations_list(for_version)	Check what AILANG can't do (Y-combinator etc.) before proposing it
effects_catalog(for_version)	Find which capability flag a function needs
submit_feedback(...)	File a bug/limitation report mid-session — lands in our triage queue

Pass for_version=$(ailang prompt --version-active) (or empty for "latest") so the response is guaranteed to match the user's CLI capabilities.

If the harness does NOT support remote MCP — fall back to the local CLI:

ailang prompt                            # Full teaching prompt (embedded)
ailang prompt --source mcp               # Force fresh copy from MCP if reachable
ailang prompt --source embedded          # Force embedded copy (eval reproducibility)
ailang docs std/list                     # Inspect a stdlib module locally
ailang examples search "pattern matching"
ailang mcp status                        # Check MCP reachability + drift

ailang prompt --source auto (the default) prefers MCP-served content when reachable AND its served_for matches the CLI version, otherwise silently falls back to embedded. Reproducible eval runs MUST set --source embedded.

Local CLI Reference (always works, even offline)

Core commands

ailang repl                                       # Interactive
ailang run --caps IO,FS --entry main file.ail     # Flags BEFORE filename!
ailang check file.ail                             # Type-check only
ailang verify file.ail                            # Z3 contract verification
ailang watch file.ail                             # Auto-reload on change
ailang iface mymodule                             # Module interface as JSON

Critical: ailang run flags MUST come before the filename. ailang run file.ail --caps IO silently ignores the flags.

Examples (already on disk)

ailang examples search "pattern matching"
ailang examples list --tags adt
ailang examples show adt_option --run

Three Module Conventions (memorize these — they bite)

1. REPL — no module declaration

ailang repl

2. Module file — standard

module myapp/main
import std/io (println)
export func main() -> () ! {IO} = println("Hello, AILANG!")

Module path uses underscores, not hyphens (mcp_tools/lib, not mcp-tools/lib).

3. Eval-harness benchmark — must use module benchmark/solution

module benchmark/solution
import std/io (println)
export func main() -> () ! {IO} = println("Benchmark solution")

Workflow

When the user asks for AILANG code:

1. Check that the agent has MCP access (if applicable): ailang mcp status or look for mcp.ailang.sunholo.com in the agent's MCP config.
2. Get current syntax reference: call prompt_get(for_version=...) via MCP, or ailang prompt locally.
3. Discover stdlib + examples before writing: call stdlib_search / examples_for_concept instead of guessing imports.
4. Write the code following the prompt's syntax. Prefer pure functions (! {}) for core logic — these are Z3-verifiable; keep effectful code (! {IO}) as thin wrappers.
5. Add contracts to pure functions when verification matters (requires/ensures). Then ailang verify solution.ail.
6. Type-check before running: ailang check solution.ail.
7. Run with the right caps: ailang run --caps IO,FS --entry main solution.ail.

Available Scripts

scripts/check_version.sh

Print the active AILANG prompt version (uses prompts/versions.json).

.claude/skills/use-ailang/scripts/check_version.sh

scripts/validate_code.sh <file.ail>

Type-check an AILANG file via ailang check.

.claude/skills/use-ailang/scripts/validate_code.sh solution.ail

Migration Note (M-AGENT-MCP)

This skill used to embed extensive syntax reference (resources/syntax_quick_ref.md, common_patterns.md, z3_verification_patterns.md). Those files are still on disk for offline use, but prefer the MCP server — it's version-locked to your CLI and updates without skill edits. See docs/guides/agent-mcp.md for the full tool catalog.

Getting Help

• AILANG MCP: https://mcp.ailang.sunholo.com/mcp/ (the canonical reference)
• Documentation: https://ailang.sunholo.com/
• Examples: ailang examples search ... or examples/ directory
• Issues: https://github.com/sunholo-data/ailang/issues — or call submit_feedback from inside an MCP-connected session
• REPL: type :help in ailang repl