Prose as syntax

The sentence is the program.

Liminate is a prose-as-syntax language designed from the human end.

filter the orders where total is above 50 is not a prompt to an AI. The interpreter reads it, works out what it means, and runs it.It is parsed, analyzed, and executed directly.

v0.14.0v0.14.0 1456 tests passing1456 pytest cases 139+ locked test sentences139+ locked test sentences 58 reserved words21 verbs, 22 connectives 4 Liminate-built agent skillscontracts + intent IR + pager + handoff Open sourceApache-2.0

Try the first program

gather the numbers from 1 to 10
filter the numbers where each is above 5
combine the numbers

Readable English. Deterministic execution.

What Liminate proves

Programs can remain close to the words people already use to describe rules, filters, conditions, and reactions.

Read why this is not prompting

The language is running

One scenario. Three products. One interpreter.

Watch a quarterly earnings review flow through the entire Liminate stack.

The scenario: An analyst uses AI to draft a Q1 earnings summary. The supervisor set the rules. The interpreter checks the work. The reviewer records what it felt like.
1 Define what right looks like Agreements
What the supervisor described
Source must be the official earnings release Revenue must match within 1% No blog posts or news articles AI drafting is fine if disclosed
The interpreter's decision
Source: earnings release
Revenue within tolerance
$24.93B vs $24.93B · delta 0.0%
No prohibited sources
AI use disclosed
✓ Passed — work meets the agreement
2 Check the AI's citations Receipts
The agent's contract
remember a source called earnings with "Revenue was $24.93 billion"
cite "24.93 billion" from earnings
measure 24.93 from earnings within 0.5
cite "guidance raised" from earnings
The receipt
cite: "24.93 billion" found
measure: 24.93 within 0.5
matched 24.93 · delta 0.0
cite: "guidance raised"
substring not found in source
✗ 1 citation failed — the AI added a claim
3 Record how the review felt Mood Ring
What the reviewer wrote
I expected to feel anxious about the earnings review, but the agreement caught the fabricated citation before I even saw it. I'm relieved. The numbers were right — it was the claim that wasn't.
The reading
Relieved
You expected anxious — you felt relieved.
Surprise: expected anxious, felt relieved — the system caught the error before you did

Small pages

This site is intentionally built as a constellation of small HTML documents. Each page explains one idea, shows one example, and points to the next page.

Agent skills

Liminate is now load-bearing outside the interpreter.

Four public Agent Skills use Liminate as their bounded, inspectable substrate: one for session truth, one for intent compilation, one for paging long context, and one for continuity handoffs.

liminate-session-contracts

Writes verified claims, open questions, locked decisions, and user corrections into a small Liminate contract.

Open repo

prosecode-prompt-compiler

Compiles user prompts into a compact seven-verb Intent IR before the agent answers.

Open repo

prosecode-context-pager

Scores long conversation blocks against current intent, then records retain, page, or evict decisions.

Open repo

prosecode-handoff-packet

Packages verified state, open questions, and corrections so another agent can continue without flattening progress into a summary.

Open repo

See how the four skills fit together · Why Liminate matters for AI agents

Try it

See Liminate in action.

Mood Ring is a color-shifting journal that runs all 58 Liminate reserved words underneath. As you write, the interpreter reads your prose and shifts the page color in real time. When you're done, you get a receipt.

Mood Ring

Write something. Watch the color change. See your emotional journey mapped by a program you can read. Share it.

OKLCH color · 21 verbs · perceptual uniformity · social share

Try Mood Ring →

The Liminate family

Prose-as-syntax beyond the interpreter.

Liminate is the foundation of the Prosecode family — a set of tools for writing, verifying, and transferring structured reasoning. The language is the substrate. These products build on it.

Receipts

An inspection surface for AI reasoning. When an agent produces a session contract, the Liminate interpreter checks its citations, measurements, and sources. The receipt shows what was verified and what wasn't.

Epistemic observability — not what agents did, but what agents believed.

receipts →

Agreements

Pre-action authority checks. Define what must be true, what is forbidden, and what is permitted before work proceeds. The interpreter enforces the rules. The receipt proves the check.

Agreements define the standard. Receipts prove the check.

agreements →

Session Contracts

Executable .limn files that track what was verified versus inferred during a working session. The contract is an artifact — inspectable, shareable, and deterministically checkable.

Open repo

Builder

Who is building this

Designed by R. Michael Thomas — working on languages that begin from how people already write, not from how machines already parse.

Liminate is one experiment in that direction. The code, examples, tests, release binaries, and build path are all open. It's well-tested: 1456 automated checks and 139+ locked test sentences so behavior doesn't drift.

Liminate is one experiment in that direction. The interpreter, examples, locked test sentences, PyInstaller build path, and release workflow are public. Schema, syntax, and behavior are pressure-tested against 1456 pytest cases and 139+ frozen sentences.

Free and open source. No framework, no platform, nothing to sign up for. If you write programs you'd rather other people could read — this is for you.

Open source under Apache-2.0. No framework, no platform, no lock-in. If you write programs you'd rather other people could read — this is for you.