> ## Documentation Index
> Fetch the complete documentation index at: https://docs.curtly.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Compression Modes

> Compare Conservative, Balanced, and Aggressive operational modes.

Curtly provides three operational modes to balance token reduction against grammatical preservation:

| Mode             | Target Savings |      Filler Removal      | Qualifier Pruning | Sentence Condensation | Table Compaction | Primary Use Case                              |
| :--------------- | :------------: | :----------------------: | :---------------: | :-------------------: | :--------------: | :-------------------------------------------- |
| **Conservative** |    20% – 35%   |   Safe (\~15 patterns)   |        None       |          None         |    Clean only    | Audited compliance, legal instructions        |
| **Balanced**     |    40% – 55%   | Standard (\~60 patterns) |      Moderate     |        Standard       |      Compact     | Production backend services, agent prompts    |
| **Aggressive**   |    55% – 75%   |        Exhaustive        |     Aggressive    |        Maximum        | Delimiter arrays | Large RAG contexts, batch document processing |

***

## 1. Conservative Mode

Focuses on non-destructive structural optimization without rephrasing:

* Normalizes consecutive blank lines and trailing whitespace.
* Drops standard conversational preambles (e.g. *"As an AI assistant..."*).
* Preserves full grammatical structure, modifiers, and adjectives.

***

## 2. Balanced Mode (Default)

Applies structural tightening and conversational condensing while maintaining clear semantics:

* Converts verbose role scaffolding into concise directives (`You are a software engineer whose task is to...` $\rightarrow$ `Role: software engineer. Task:...`).
* Drops conversational padding (`Please note that`, `It is important to remember that`).
* Prunes common low-entropy adverbs while preserving high-entropy logic operators (`never`, `must`, `always`, `only`).

***

## 3. Aggressive Mode

Maximizes token efficiency for high-volume context windows:

* Removes transition words, non-critical conjunctions, and weak qualifiers.
* Deduplicates repetitive instructions using token-set overlap analysis.
* Compresses wide Markdown tables into compact delimiter formats.