leap

KO | EN

gitlite — search

#javascript #python #hacktoberfest #ai #llm #golang #machine-learning #rust #deep-learning #windows #cli #mcp #openai #open-source

(★ 17)

LEAP — LLM Engineered Application Pattern. Code is a commodity, guardrails are the product.

leap Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# LEAP

> **L**LM **E**ngineered **A**pplication **P**attern
>
> A development philosophy where **code is a commodity** and **guardrails are the product**.

## The pitch

Stop writing code. Write the specification.

A LEAP project contains no application code. Instead, it contains:

- **Prompts** — declarative intent
- **Schemas** — contracts between components
- **Tests** — behavioral guardrails

Any AI agent can generate the code on demand. The code is gitignored. The prompts, schemas, and tests are the source of truth.

## Proof that this works

We reimplemented [SQLite](https://www.sqlite.org/) from one language-neutral specification into **five** languages at once:

| | [sqlite-leap](https://github.com/safitudo/sqlite-leap) | mainline SQLite |
|---|---|---|
| Targets | **C, Rust, Zig, Go, Python** | C |
| Source LOC (5 targets total) | ~250,000 | ~150,000 |
| Spec LOC | **~28,000** | n/a (the code is the spec) |
| Upstream sqllogictest excl-SKIP | **98.88–99.98%** per target | 100% |
| `exec/mainline` (denominator parity) | 99.93–99.97% on 4 compiled targets | 100% |
| Crashes on full corpus | **0** across 4 compiled targets | 0 |
| On-disk byte-identity at fixed fixtures | **5/5 + mainline match** (SHA1) | n/a |

One language-neutral spec, five languages, byte-identical `.db` files at fixed fixtures, full upstream test corpus. Methodology in [MULTI_TARGET.md](MULTI_TARGET.md).

We also did this earlier with [npm's semver parser](https://github.com/npm/node-semver) in a single language, as the original LEAP proof:

| | [semver-leap](https://github.com/safitudo/semver-leap) | node-semver |
|---|---|---|
| Source LOC | 2,540 | 2,271 |
| Tests passing | **5,632 / 5,632** | 5,632 / 5,632 |
| Development time | **~1 session** | 15 years |
| Contributors | 1 human + AI | 106 humans |
| Human-authored lines | **717** (prompts + schemas) | 2,271 (code) |

One session. 717 lines of spec. A production-grade semver parser.

## The thesis

The internet is benchmarking AI on "how well does it write code."

That's the wrong question.

**Code is a commodity.** Well-specified libraries aren't novel algorithms — they're 15 years of "fix edge case X, add test for Y." Those edge cases are already in tests. If they're in tests, AI can eat them.

The real questions are:

1. How good are your **schemas** at defining boundaries?
2. How good are your **tests** at catching when generated code is wrong?
3. How good are your **prompts** at conveying intent?

Value shifts from **writing code** to **specifying behavior**. Tests aren't a safety net — they're the product.

## The Guardrails Hierarchy

Three artifacts, ranked by value:

1. **Tests** — most valuable. Human-authored. The definition of correctness.
2. **Schemas** — the architecture. Contracts between parts.
3. **Prompts** — least valuable. Disposable. Any prompt that passes the tests is valid.

If you lose a prompt but keep the tests and schemas, any new prompt that passes is equally correct.

## Quick start

### 1. Install the LEAP skill (Claude Code)

```
/plugin marketplace add safitudo/leap-skill
/plugin install leap-skill@leap-skill
```

### 2. Use it on a LEAP project

```bash
git clone https://github.com/safitudo/semver-leap.git
cd semver-leap
npm install

# In Claude Code:
claude "build the app, start from master.md"
```

### 3. Or start your own

Read [SPEC.md](SPEC.md) for the project structure.
Read [MANIFESTO.md](MANIFESTO.md) for the philosophy.
Read [AGENTS.md](AGENTS.md) if you use a non-Claude agent.
Read [PIXEL_PERFECT.md](PIXEL_PERFECT.md) if you're building UI.
Read [MULTI_TARGET.md](MULTI_TARGET.md) if you're targeting multiple languages from one spec.

## Examples

- **[sqlite-leap](https://github.com/safitudo/sqlite-leap)** — SQLite, 5 targets (C, Rust, Zig, Go, Python) from one language-neutral spec, 99.93–99.97% denominator parity on the upstream sqllogictest corpus
- **[semver-leap](https://github.com/safitudo/semver-leap)** — npm semver parser, 5,632/5,632 tests passing
- **[todolist-leap](https://github.com/safitudo/todolist-leap)** — the original proof of concept

More proofs coming daily. Follow [@safitudo](https://github.com/safitudo) for updates.

## Is this for real?

Yes. The semver-leap repo is a real working package that passes the upstream test suite verbatim. You can run `npm test` yourself.

## The bigger bet

We think most open source libraries can be generated this way. Not because AI is magic, but because:

- Most libraries aren't novel algorithms
- 15 years of bug fixes is mostly captured in test suites
- Test suites ARE the spec for "correct behavior"
- If the spec is complete, the implementation is mechanical

LEAP is betting that **the spec is the thing**. Code is what falls out.

## License

MIT

## Roadmap & contributing

- **[ROADMAP.md](./ROADMAP.md)** — what's proven, what's in flight, and the unsolved front (UI / pixel-perfect). Open invitations for library stunts, cross-model verification, and UI experiments.
- **[CONTRIBUTING.md](./CONTRIBUTING.md)** — how to propose SPEC/MANIFESTO edits, submit examples, or port another library as a stunt.
- **[Discussions](https://github.com/safitudo/leap/discussions)** — open-ended questions, counter-arguments, "what if LEAP worked like X?"
- **Issues** — actionable proposals and bugs in the methodology.

## Home

[leapagentic.io](https://leapagentic.io) — the home of LEAP.

leap

leap Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

leap

leap Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README