How Skimmy Works

Skimmy uses a unique hybrid architecture combining a deterministic grammatical engine with a lightweight local Large Language Model (like Gemini Nano).

1. The Tripartite Argument Profile (inspired by Stephen Toulmin)

When you open an article, Skimmy first samples the introduction, middle, and conclusion. It securely passes this cross-section to your local LLM to generate a Tripartite Argument Profile.

Instead of reading every single word, the LLM simply extracts three key lists:

• Status Quo Entities: The problem, competitors, legacy tools, or concepts the author is arguing against.
• Core Argument Entities: The proposed solution, new features, or concepts the author is arguing for.
• Action Verbs: The verbs driving the narrative forward.

2. The English Grammar Engine

With the profile in hand, Skimmy's deterministic grammar engine takes over. It parses the grammatical tree of the entire article locally in your browser.

At the sentence level, Skimmy guarantees readability by targeting specific syntactic constructs using a mathematical formula:

• Lexical Density & Fluff Penalties: Sentences that fall below a mathematically optimized "signal-to-noise" ratio are struck. Clauses that lack core argument entities or are entirely comprised of "status quo" background context are crossed out.
• Structural Protections: Skimmy automatically protects proper nouns, direct quotes, exact data phrases (numbers, percentages), and clauses containing critical action verbs.
• Rambling Prevention: If an author begins to ramble (e.g., three consecutive fluffy clauses), the engine dynamically tightens its density thresholds, punishing elaboration.

3. Auto-Research Optimization (inspired by Andrej Karpathy)

How do we know which grammatical rules actually work? We don't guess.

Skimmy was built using an autonomous Auto-Research Optimizer. We created a gold-standard dataset of highly complex texts—from dense Stratechery tech analysis to fast-paced news to jargon-heavy thought leadership.

Our automated engine brute-forced thousands of parameter combinations (testing different density thresholds, proper noun bonuses, and clause lengths) to find the mathematically perfect configuration that maximizes signal retention while aggressively striking fluff.

Non-English Fallback

Because our grammatical engine is deeply tuned for English syntax rules, it won't work on languages like Spanish, Japanese, or Korean.

To solve this, Skimmy automatically detects the language of the page you are on. If it detects a non-English language, it bypasses the English grammatical shields and falls back to a purely semantic summarization model using your local LLM.