# How Humanizer Decides to Cut or Keep Claims When Applying Inflation Patterns

> Learn how Humanizer decides to cut or keep claims when applying inflation patterns. Get a lean, fact-preserving rewrite by removing exaggeration and vague phrasing.

- Repository: [Siqi Chen/humanizer](https://github.com/blader/humanizer)
- Tags: deep-dive
- Published: 2026-09-12

---

**Humanizer keeps claims that are directly supported by verifiable source material and cuts any language that merely inflates significance, adds unwarranted authority, or uses vague sales-style phrasing, resulting in a lean, fact-preserving rewrite.**

The `blader/humanizer` repository implements a systematic approach to detect and remove "inflation" from text—excessive language that exaggerates importance or invokes borrowed authority without substantiation. When determining whether to cut or keep specific claims during inflation pattern editing, Humanizer evaluates each statement against its underlying source support, stripping only the rhetorical dressing while preserving the factual core.

## The Six-Step Decision Framework

According to the source code in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) (lines 194-275), Humanizer treats inflation as "dressing" around an otherwise sound fact. The algorithm follows six distinct steps when processing section **C. Inflation and borrowed authority**:

### Step 1: Isolate the Verifiable Core

Humanizer first identifies the concrete claim that can be verified from the source material. This includes specific dates, events, statistics, or observable details that form the factual foundation of the sentence.

### Step 2: Validate Direct Source Support

If the underlying fact is directly supported by the source, the claim is **kept**. Any assertion that cannot be traced back to the provided source material is flagged for removal.

### Step 3: Strip Inflated Significance

Phrases that add pomp without substance—such as "stands as a testament," "pivotal moment," or "future looks bright"—are removed, leaving only the factual core. This *Inflated significance* rule is defined in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 194-222.

### Step 4: Resolve Vague Connections

Claims that merely associate two items without explaining the concrete relationship (e.g., "associated with," "linked to") are either reduced to the specific connection or cut entirely if the relationship remains unclear. This logic appears in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 224-232.

### Step 5: Eliminate Borrowed Authority Fluff

Overused AI buzzwords, sales language, and unnamed expert citations are stripped unless the source explicitly provides the needed authority. Humanizer removes this fluff according to the rules in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 198-214 and 250-257.

### Step 6: Normalize Qualified Statements

Qualifiers such as "may," "possibly," or "might" are retained only when the source material explicitly conveys that level of uncertainty. Otherwise, these hedges are simplified to definitive statements per [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 268-275.

## Real-World Editing Examples

The following examples from the `blader/humanizer` codebase demonstrate how the algorithm parses sentences, isolates factual cores, and removes inflated language while determining which claims to preserve.

### Example 1: Inflated Significance

```text
Original: "The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain."

Humanized: "The Statistical Institute of Catalonia was established in 1989, part of a wider decentralization of administrative functions in Spain."

```

### Example 2: Vague Connection

```text
Original: "He is associated with the Rajhans Orchestra, which he founded and conducts."

Humanized: "He founded and conducts the Rajhans Orchestra."

```

### Example 3: Borrowed Authority

```text
Original: "Her views have been cited in The New York Times, BBC, Financial Times, and The Hindu."

Humanized: "Her views have been cited in The New York Times and the BBC."

```

## Source File Reference

The inflation pattern implementation resides in specific documentation and source files:

- [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) – Contains the complete definition of all inflation patterns and the decision rules for cutting versus keeping claims (lines 194-275)
- [`README.md`](https://github.com/blader/humanizer/blob/main/README.md) – Provides the high-level overview of the inflation section and user-facing examples (lines 102-110)

## Summary

- Humanizer evaluates claims against direct source support to determine whether to cut or keep specific claims when applying inflation patterns
- Inflated significance markers like "pivotal moment" are stripped according to [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 194-222
- Vague associative language is resolved to concrete relationships or removed to prevent unsubstantiated connections
- Borrowed authority and sales-style fluff are deleted unless explicitly sourced in the original material
- Qualified uncertainty is preserved only when the source material warrants it, per [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 268-275

## Frequently Asked Questions

### What qualifies as an "inflation pattern" in Humanizer?

Inflation patterns include significance-inflating phrases, vague connections between entities, and borrowed authority claims that lack specific source attribution. These are defined in section C of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) as language that adds "dressing" around otherwise sound facts without providing verifiable substance.

### How does Humanizer handle uncertain or qualified statements?

Qualifiers like "may" or "possibly" are kept only when the source material explicitly conveys that level of uncertainty. Otherwise, Humanizer simplifies the statement to a definitive claim, following the normalization rules specified in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 268-275.

### Can Humanizer distinguish between legitimate authority and borrowed authority fluff?

Yes. Humanizer cuts unnamed expert citations and overused AI buzzwords unless the source explicitly provides the authority. This distinction is enforced through the rules documented in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 198-214 and 250-257, which target sales-style language and unattributed expert claims.

### Where are the inflation pattern rules documented?

The complete decision framework is documented in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) lines 194-275, with a high-level overview and usage examples available in [`README.md`](https://github.com/blader/humanizer/blob/main/README.md) lines 102-110 at the `blader/humanizer` repository.