# How Humanizer Uses Built-In Safeguards to Prevent Invention of Factual Information > Discover how Humanizer's four layered safeguards prevent factual invention in rewrites. Learn about its absolute ban on making things up, pattern stripping, validation, and user prompts. - Repository: [Siqi Chen/humanizer](https://github.com/blader/humanizer) - Tags: how-to-guide - Published: 2026-09-12 --- **Humanizer prevents factual invention through four layered defenses codified in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md): an absolute ban on making things up, specific patterns that strip unsupported claims, validation that preserves every original fact, and user prompts when information is missing.** The **blader/humanizer** repository implements a zero-tolerance policy for hallucinated details during text rewriting. Its built-in safeguards prevent invention of factual information by embedding strict instructions directly into the skill's prompt architecture rather than relying on executable validation code. These protections operate through carefully engineered markdown instructions in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md), [`README.md`](https://github.com/blader/humanizer/blob/main/README.md), and [`AGENTS.md`](https://github.com/blader/humanizer/blob/main/AGENTS.md) that govern how the language model processes and transforms text. ## The Core "Do Not Make Anything Up" Rule At line 73 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md), the skill declares its foundational constraint: **"Do not make anything up."** This instruction mandates that every name, number, date, quote, citation, or factual detail must originate from the source text or the writer. When a rewrite requires a detail that does not exist in the input, Humanizer must ask the user rather than generate the information, effectively blocking all hallucination at the instruction level. ## Pattern-Based Safeguards for Unsupported Claims Beyond the core rule, Humanizer employs specific patterns to catch and eliminate different categories of unsupported assertions before they reach the final output. ### Pattern 23: Knowledge-Limit Disclosures and Guesses Lines 25-33 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) define **Pattern 23**, which targets sentences attempting to fill knowledge gaps with speculation. When the source lacks specific information, this pattern either removes the speculative sentence entirely or rewrites it to explicitly state that the source does not provide the information. This prevents the model from guessing dates, figures, or outcomes when the original text remains silent on those details. ### Pattern 17: Borrowed Authority Without Sources Lines 53-55 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) address **Pattern 17**, which handles citations to unnamed authorities. When text attributes claims to vague "experts" or "studies" without specific sources, Humanizer either preserves the original phrasing exactly or removes the unsupported attribution. The skill never fabricates a source name, publication, or study to make the citation appear concrete. ## Multi-Tell Validation and Fact Retention The **multi-tell safeguard** treats facts as guard rails that must survive the rewriting process intact. According to lines 36-38 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) and lines 73-74 of [`README.md`](https://github.com/blader/humanizer/blob/main/README.md), the validator checks that the final draft contains every supported claim present in the original. Any loss of a factual claim during rewriting triggers an error state, forcing the model to maintain factual fidelity throughout the transformation process. ## Voice Matching Without Factual Extrapolation When users provide personal writing samples, Humanizer mimics rhythm and word choice but strictly forbids extrapolating facts from the sample into new contexts. As documented in lines 42-44 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) and lines 64-66 of [`README.md`](https://github.com/blader/humanizer/blob/main/README.md), the skill copies stylistic elements only where the sample already contains the specific factual information, never assuming or inventing details beyond what the user explicitly provided. ## Examples: Safeguards in Action The following demonstrations show how these rules execute when invoking the skill via the CLI (`/humanizer`) or Claude plugin. **Example 1: Removing Invented Hardware Specifications** ```text /humanizer The server processes 10,000 requests per second, a speed only achievable on the upcoming "Nimbus X" hardware. ``` **Humanizer's Response:** ```text The server processes 10,000 requests per second. ``` The clause citing "Nimbus X" hardware invents a specific product name not present in the source. Pattern 23 triggers removal of the unsupported specification while preserving the verified throughput metric. **Example 2: Eliminating Date Speculation** ```text /humanizer While the exact launch date isn't public, it is likely to be early 2024. ``` **Humanizer's Response:** ```text The exact launch date isn't public. ``` Pattern 23 removes the speculative "likely to be early 2024" because the source confirms only the absence of the date, not the probable timing. **Example 3: Stripping Unattributed Authority** ```text /humanizer Experts believe the algorithm will double user engagement. ``` **Humanizer's Response:** ```text The algorithm will double user engagement. ``` Pattern 17 removes the "Experts believe" attribution because no specific experts or studies are cited in the source, preventing the invention of authority. ## Summary - The **"Do not make anything up"** rule in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) (line 73) establishes an absolute prohibition against inventing names, dates, numbers, or citations. - **Pattern 23** (lines 25-33) eliminates speculative guesses about unknown information, replacing them with explicit statements about knowledge limits. - **Pattern 17** (lines 53-55) removes vague attributions to unnamed authorities rather than fabricating specific sources. - **Multi-tell validation** (lines 36-38) verifies that every factual claim from the source survives the rewrite process without omission. - **Voice matching** (lines 42-44) confines stylistic imitation to existing facts without extrapolating new information from user writing samples. - **Cross-platform consistency** is enforced through [`AGENTS.md`](https://github.com/blader/humanizer/blob/main/AGENTS.md), ensuring these safeguards operate identically across Claude, OpenAI, and other supported agents. ## Frequently Asked Questions ### What happens if Humanizer needs a fact that isn't in the source text? According to line 73 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md), Humanizer must ask the user for the missing detail rather than inventing it. The skill treats any required factual element—whether a name, number, date, or citation—as something that must come from either the source material or direct user input, with no exceptions for plausible-sounding fabrications. ### Does Humanizer ever create fake citations to support claims? No. Pattern 17 in [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md) (lines 53-55) explicitly forbids inventing sources. When encountering borrowed authority without specific attribution, Humanizer either keeps the vague phrasing intact or removes the attribution entirely, but never fabricates a source name, study, or publication to make the claim appear verified. ### How does Humanizer handle speculation about future dates or outcomes? Pattern 23 (lines 25-33 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md)) catches these knowledge-limit disclosures automatically. The pattern either deletes the speculative sentence or rewrites it to state that the information is not publicly available, preventing the model from guessing release dates, performance metrics, or other unverified predictions. ### Can Humanizer add facts from my writing samples to new content? No. While the skill can match your voice using personal samples (lines 42-44 of [`SKILL.md`](https://github.com/blader/humanizer/blob/main/SKILL.md)), it only borrows stylistic elements where the sample already contains the specific factual information. It will not extrapolate facts from your sample into new contexts or invent details based on patterns observed in your writing style.