CompactPrompt

CompactPrompt shortens the text you send to an AI model while preserving its meaning. The result costs less to run, returns faster, and is less likely to exceed the model's input limit. The common case is a single function call, and no background in machine learning is required to use it.

Background

An AI model reads an input — the prompt — and returns a response. Providers charge according to the amount of text processed, measured in tokens (each token is roughly three-quarters of a word), and every model has a maximum input size. A long prompt that combines instructions, documents, tables, and examples therefore costs more, responds more slowly, and may not fit at all.

CompactPrompt reduces the size of a prompt while retaining the information that matters, so you keep the substance and discard the overhead.

Getting started

Install the library:

pip install compactprompt

Shorten a prompt:

from compactprompt import CompactPrompt

result = CompactPrompt.compact(
    "Please could you very kindly go ahead and provide a really concise "
    "summary of the quarterly report."
)

print(result.text)
print(f"{result.ratio:.1f}x smaller "
      f"({result.tokens_before} -> {result.tokens_after} tokens)")

Output:

a really concise summary of the quarterly report.
1.7x smaller (22 -> 13 tokens)

The filler — "Please could you very kindly go ahead and" — is removed, and the meaning is unchanged.

What it does

CompactPrompt provides several methods for reducing the size of a prompt. They can be used individually or together. Each is described below in plain terms, followed by an optional technical note.

Trimming low-value wording

Removes words that carry little meaning, such as conversational filler, and keeps the words that do. This is lossy: the removed words are not recoverable, but the result is ready to use as it is.

from compactprompt import CompactPrompt

# Remove approximately 40% of the tokens
result = CompactPrompt.compact(prompt, ratio=0.4)

# Or target a specific size
result = CompactPrompt.compact(prompt, budget=64)

<details> <summary>Technical detail</summary>

Each word receives an information score that combines how rare it is in general (static self-information) with how predictable it is in context (dynamic self-information from a small language model). Low-scoring words are removed. Whole grammatical phrases are removed together, using spaCy, so the result remains readable, and names and numbers are protected. This implements the fusion rule from the CompactPrompt paper. </details>

Reversible shortening of repeated phrases

When a phrase recurs, it is replaced with a short placeholder, and a key records what each placeholder stands for. This is lossless: the exact original can be restored at any time.

import compactprompt as cp

doc = "operating cash flow rose. operating cash flow fell. operating cash flow held."
abbr = cp.abbreviate(doc, n=3)

print(abbr.text)        # '@0 rose. @0 fell. @0 held.'
print(abbr.dictionary)  # {'@0': 'operating cash flow'}
print(abbr.restore())   # the exact original

Retain abbr.dictionary so the placeholders can be expanded again later.

Reducing the size of numeric data

Large tables of numbers consume many tokens. This lowers their precision to save space while guaranteeing that the rounding never exceeds a known bound.

import compactprompt as cp

q = cp.quantize([1.0, 2.5, 3.3, 4.8, 9.2, 10.0], bits=8)
q.reconstruct()   # the rounded values
q.max_error       # the guaranteed maximum error

Selecting representative examples

Models perform better when shown a few examples. If you have many candidate examples, this selects a small, varied subset that still reflects the full range, so you send a representative few rather than all of them.

from compactprompt import select_examples

chosen = select_examples(my_examples)
chosen.examples

Choosing how the wording is trimmed

The wording-trimming step can be carried out by any of three interchangeable engines. All of them shorten text; they differ in how they decide what to remove and in what they require to run. Select one with the engine argument — nothing else in your code changes.

Engine	Approach	Requirements
Built-in (default)	Scores each word and removes the least useful. Runs locally.	None
LLMLingua	Microsoft's established tool, which uses a small model to decide what to remove.	Downloads a model
Caveman	Rewrites the text in a concise style, preserving code, links, and headings.	Access to a language model

CompactPrompt.compact(prompt)                       # built-in, no extra install
CompactPrompt.compact(prompt, engine="llmlingua")   # pip install 'compactprompt[llmlingua]'
CompactPrompt.compact(prompt, engine="caveman")     # pip install 'compactprompt[caveman]'

The built-in engine and the other core features implement the CompactPrompt research paper. LLMLingua and Caveman are independent open-source tools that this library integrates; see Attribution.

Compacting files and skills

CompactPrompt can also compact whole markdown files — documentation, CLAUDE.md, notes, and Claude Code skills (SKILL.md) — not just strings. It can first review a file or folder to report where the savings are.

This works safely by design: YAML frontmatter is preserved exactly, fenced code blocks and links are never altered, the result is rejected if it would change a heading, code block, or URL, and nothing is written without --apply (which first saves a .bak backup). Files that look like code, config, or secrets are skipped automatically.

From the command line:

# See where the savings are (read-only)
compactprompt review ./skills

# Preview the compaction of one skill (writes nothing)
compactprompt compact ./skills/my-skill/SKILL.md --engine builtin

# Apply it (saves SKILL.md.bak, then rewrites the file)
compactprompt compact ./skills/my-skill/SKILL.md --engine caveman --apply

--engine is required — you choose builtin, llmlingua, or caveman each time (caveman, which rewrites prose, is usually best for human-readable files).

From Python:

from compactprompt import review_file, compact_file

report = review_file("SKILL.md")
print(report.tokens, report.issues)

result = compact_file("SKILL.md", engine="caveman", apply=True)
print(result.tokens_before, "->", result.tokens_after)

The Streamlit app's Files & Skills tab does the same interactively.

Optional features

The basic installation requires no setup. Additional features depend on extra components, which you install only as needed:

pip install compactprompt                  # core: trimming and reversible shortening
pip install 'compactprompt[ml]'            # numeric reduction and example selection
pip install 'compactprompt[llmlingua]'     # the LLMLingua engine
pip install 'compactprompt[caveman]'       # the Caveman engine
pip install 'compactprompt[mcp]'           # the MCP server for AI agents
pip install 'compactprompt[app]'           # the interactive application
pip install 'compactprompt[all]'           # everything

When a feature needs a component that is not installed, CompactPrompt reports exactly what to install.

Interactive application

A small web application lets you paste a prompt and see it shortened, with the savings reported as you go:

pip install 'compactprompt[app]'
streamlit run compactprompt_app.py

It opens in the browser. Use the sidebar to choose an engine and set how much to remove.

Use it from an AI agent

CompactPrompt ships an MCP server so AI coding tools (Claude Code, Codex, Cursor, Gemini, and any MCP-capable agent) can review and compact prompts, docs, and skills directly:

pip install 'compactprompt[mcp]'      # provides the `compactprompt-mcp` command
claude mcp add compactprompt -- compactprompt-mcp   # e.g. for Claude Code

The agent-skills/ directory also has lightweight skill/rules files and an install.sh for the same tools. See its README for per-tool configuration.

<!-- mcp-name: io.github.gtkcyber/compactprompt -->

Confirming the meaning is preserved

To check that a shortened prompt still means the same thing, you can measure the similarity between the original and the result, where 1.0 indicates identical meaning:

from compactprompt import cosine_fidelity   # pip install 'compactprompt[embeddings]'

score = cosine_fidelity(original_text, result.text)
print(score.mean)

Reference

CompactPrompt.compact(...) returns a result object with the following fields:

Field	Meaning
.text	The shortened prompt.
.original	The input.
.tokens_before / .tokens_after	Size before and after.
.ratio	How many times smaller (for example, 2.3).
.savings	Fraction of tokens saved (for example, 0.4).
.dictionary	The key for restoring shortened phrases, when used.
.restore()	Reverses the reversible shortening step.

The principal options:

CompactPrompt.compact(
    prompt,
    ratio=0.5,          # how much to remove: 0.5 targets about half the tokens
    budget=None,        # alternatively, a specific target token count
    prune=True,         # trim the wording (default)
    abbreviate=False,   # also shorten repeated phrases (reversible)
    engine="builtin",   # "builtin", "llmlingua", or "caveman"
)

The complete reference, including the advanced options, is in the documentation.

Development

Run the tests:

pip install pytest
pytest

The suite runs against the dependency-free core; tests for optional features are skipped automatically when those components are absent.

Build the documentation locally:

pip install 'compactprompt[docs]'
mkdocs serve

Citation

This library implements the methodology from:

@article{choi2025compactprompt,
  title={CompactPrompt: A Unified Pipeline for Prompt and Data Compression in LLM Workflows},
  author={Choi, Joong Ho and Zhao, Jiayang and Shah, Jeel and Sonawane, Ritvika and
          Singh, Vedant and Appalla, Avani and Flanagan, Will and Condessa, Filipe},
  journal={arXiv preprint arXiv:2510.18043},
  year={2025}
}

It is an independent implementation and is not affiliated with the authors of the paper.

Attribution

The Caveman engine (compactprompt/caveman.py) is adapted from Caveman by Julius Brussee (MIT). The LLMLingua engine uses LLMLingua by Microsoft (MIT). Full third-party attributions and license notices are in THIRD_PARTY_NOTICES.md.