A gallery of real programs built in Promise by AI agents — each with the exact prompt, the model that wrote it, the language version, and an honest account of how the run actually went.
Promise is a statically-typed, natively-compiled language designed so an AI agent can write correct, maintainable code — and whose own compiler and standard library are themselves written by AI agents.
This repo is a record of that in practice. Each entry is one run: hand an agent a prompt, have it build something in Promise, and record what happened — the prompt, the agent and model, the Promise version, how many iterations / how long it took, where it got stuck, and the resulting project.
It's meant to be read for yourself, not a benchmark — and it keeps the rough edges, not just the clean wins.
Each task is a folder with its own prompt.md (the task-specific ask) and a
README.md results table across agents (see TEMPLATE.md). The
prompt actually sent to each agent wraps that prompt.md with the repo-root
PROMPT_PREFIX.md (shared "learn Promise first" preamble) and
PROMPT_SUFFIX.md (shared "write a SUMMARY, and report any
compiler bug or missing-library gap upstream" postamble). Each
agent's run lands in a <task>-<agent>/ subdir (e.g.
hello-world/hello-world-claude/) — named so Promise's folder-derived binary is
descriptive (./hello-world-claude) — containing the generated .pr source,
SUMMARY.md (the agent's own TL;DR of how it went), context.md (provenance), and
the recording (demo.cast, viewed with the asciinema player
— asciinema play demo.cast). Runs are captured with bin/record.sh.
A run that trips over a reproducible compiler bug, or needs something Promise doesn't
have yet, also leaves a BUG-*.md or FEATURE-*.md in its subdir — so each entry is
not just "did it build" but feedback on what to fix, what library module to add, or how
Promise could implement the prompt better.
- Non-deterministic. Agents don't produce identical output twice — each entry is "what happened in this run," not "run this and reproduce it exactly."
- Version-pinned. Every run records the Promise version it was built against; the language is evolving fast, so older runs may not build on newer epochs.
- Early. Promise is under active development and not production-ready.
Some tasks are inspired by work done elsewhere; where they are, the task's own
README.md gives full credit and links the source. In particular,
mini-git/ is inspired by the MiniGit exercise from
ai-coding-lang-bench by mame
(Yusuke Endoh). The zoo's prompts are written from scratch — we don't reproduce
others' task specs or test suites — and because the language, harness, and run
conditions all differ, zoo runs are never comparable to the benchmarks that
inspired them.
Dual-licensed Apache-2.0 OR MIT (at your option), matching the Promise project — see LICENSE-APACHE, LICENSE-MIT, and NOTICE.
Maintained by Promise Lang LLC. Learn about the language at promise-lang.org. Questions or want an early look? early@promise-lang.org.