MumbleLang

MumbleLang is an experiment in prompt compression.

It turns messy human requests into a compact, structured form that an LLM can still understand. The goal is to preserve the meaning, constraints, tone, and important exact phrases while removing extra language.

This repo includes MumbleLang Lab, a small local web app for testing the idea.

Why this exists

Human language is full of repetition, hedging, tone, implied constraints, and half-structured intent. LLMs can usually understand it, but they also have to process all of it.

MumbleLang asks a narrower question:

Can we convert human input into a smaller intermediate form without losing what matters?

The current answer is: sometimes. This app helps test where it works, where it saves space, and where it breaks.

What it does

Paste in a prompt, note, email, meeting transcript, customer message, technical plan, or rough idea. MumbleLang Lab converts it into one of two formats.

Strict MBL/1.0

Strict mode uses a small set of tags, including:

@task
@tone
@ctx
@goal
@build
@defer
@requires
@guard
@metric
@residue

Strict mode is best for technical work, planning, specs, legal-ish language, invention notes, and cases where scope matters.

MBL-Micro

Micro mode is shorter and looser. It is easier to scan, but it has a higher risk of losing nuance.

Use Micro for quick summaries and Strict mode when accuracy matters.

What the app checks

MumbleLang Lab reports:

• how much shorter the converted text is
• whether the output stayed inside the selected format budget
• whether exact quotes or special phrases were preserved
• whether the input is vague, emotional, legal-sensitive, or likely to drift
• whether the request depends on missing facts

The missing-facts check matters. If the source text asks for information that is not present, the app should flag that instead of inventing an answer.

Run locally

Requires Python 3.10+.

git clone https://github.com/RegularJoe-CEO/MumbleLang.git
cd MumbleLang
python app.py

Open:

http://localhost:8000

To use a different port:

PORT=3000 python app.py

Configure a model

MumbleLang Lab uses an OpenAI-compatible chat completions API.

Set these environment variables:

export LLM_API_KEY="your_api_key"
export LLM_BASE_URL="https://api.openai.com/v1"
export LLM_MODEL="gpt-4.1-mini"

Example for xAI/Grok-compatible endpoints:

export LLM_API_KEY="your_api_key"
export LLM_BASE_URL="https://api.x.ai/v1"
export LLM_MODEL="grok-4"

Do not commit real API keys.

API

Convert text

POST /api/convert

{
  "input": "Paste human text here",
  "mode": "strict"
}

Draft an answer from MumbleLang

POST /api/answer

{
  "mumblelang": "MBL/1.0\n@task ...",
  "mode": "strict"
}

Current limitations

This is an early lab build.

• Reduction is estimated by character count, not tokenizer-specific token count.
• The converter currently uses an LLM, not a deterministic parser.
• Short or vague inputs may not compress well.
• Emotional tone and personal voice are harder to preserve than technical structure.
• This is an experimental tool, not a source of legal, medical, or financial advice.

Project status

MumbleLang is not a standard. It is a working notation and test harness.

The next useful work is to collect examples, compare outputs across models, and identify which kinds of language survive compression cleanly.

License

MIT