What Is the No Invention Rule in Humanizer? Preventing AI Hallucinations at the Source

Humanizer's "no invention" rule is a core safeguard that prohibits the skill from generating facts, numbers, names, dates, or quotes not explicitly present in the source material or supplied by the user.

The blader/humanizer repository provides a Claude skill designed to rewrite AI-generated text into natural human prose. Central to its architecture is a strict editorial policy that eliminates hallucinations by refusing to fabricate information, ensuring every output remains traceable to verifiable input.

How the No Invention Rule Works

The rule operates through three enforcement layers embedded in the skill's workflow.

Preserving Existing Facts

When processing text, Humanizer may shorten, merge, or restructure sentences, but the factual content must remain unchanged. In SKILL.md, the draft-stage instructions explicitly mandate that the skill keeps every supported claim intact. If a sentence requires a detail absent from the source, Humanizer does not interpolate; instead, it halts and requests clarification from the user. This behavior is codified in the Draft rewrite section at line 36 of SKILL.md.

Explicit Prohibition of Fabricated Details

The README.md at line 73 reinforces this constraint in user-facing documentation, stating that Humanizer "does not make things up." Specifically, any name, number, date, quote, citation, or factual detail must originate from either the source text or direct user input. When encountering a gap in required information, the skill responds with a query rather than an invention, preventing the propagation of fabricated data downstream to consumers like Claude or OpenAI models.

Versioned Enforcement

The rule gained formal status in version 2.9.0, documented in the Version history section at line 181 of README.md. This versioning ensures that downstream systems can depend on consistent, fact-accurate behavior, as the safeguard is treated as a contractual guarantee rather than a loose guideline.

Practical Examples of the Rule in Action

The following examples demonstrate how Humanizer handles complete versus incomplete source material.

Preserving Provided Facts

When all necessary details are present, Humanizer restructures while maintaining accuracy:

/humanizer
[Paste your AI-generated text here]

# Example input

"The new API will launch next quarter, and it will support 10,000 concurrent users."

# Humanizer’s output (no invented detail)

"The new API will launch next quarter and will support 10,000 concurrent users."

Requesting Clarification on Missing Data

When a specific fact is missing, the skill refuses to hallucinate a value:

/humanizer
[Paste your AI-generated text here]

# Example input lacking a needed fact

"The new API will launch next quarter and will support millions of users."

# Humanizer’s response (asks for clarification)

"I need clarification: how many concurrent users does the API support? The source does not provide that number."

Installation and Invocation

To deploy the safeguarded skill locally:

# Using the Skills CLI to install Humanizer

npx skills add blader/humanizer --global

# Then invoke the skill via the agent

/humanizer

Where the Rule Is Enforced

The no invention constraint is distributed across multiple files to ensure validation at different stages:

• SKILL.md – Contains the authoritative prompt including the no-invention instruction in the Draft rewrite step.
• README.md – Documents the rule for users and tracks its formalization in the version history.
• .claude-plugin/plugin.json – Directs Claude agents to the skill, indirectly enforcing the rule for Claude-based usage.
• scripts/validate-package.py – Validates that skill files are correctly structured before publication, ensuring the rule remains intact in distributed versions.

Summary

• The no invention rule prevents Humanizer from adding names, numbers, dates, or quotes not found in the source material.
• Missing information triggers clarification requests rather than hallucinations, protecting downstream systems from fabricated data.
• The rule is hardcoded in SKILL.md and documented in README.md, with validation logic in scripts/validate-package.py.
• Version 2.9.0+ treats this safeguard as a versioned contractual guarantee for all consumers.

Frequently Asked Questions

What happens if Humanizer encounters a statistic missing from the source text?

When Humanizer requires a specific number, percentage, or metric absent from the input, it will output a request for clarification such as "I need clarification: [specific detail]? The source does not provide that number." This prevents the interpolation of plausible-sounding but fabricated statistics.

Does the no invention rule apply to stylistic rephrasing?

No. The rule specifically targets factual claims, not writing style. Humanizer may alter sentence structure, vocabulary, and tone freely, provided that every factual assertion in the original text remains intact and no new factual assertions are introduced.

How do I verify that my installed version includes the no invention safeguard?

Check the version tag in your installation against the README.md version history. The rule was formally codified as "the rule against invented facts" in version 2.9.0. You can also inspect SKILL.md line 36 to confirm the Draft rewrite instruction contains the prohibition against inventing facts.

Can the rule be disabled or overridden in the configuration?

No. According to the source code in SKILL.md and the validation logic in scripts/validate-package.py, the no invention constraint is baked into the core skill instructions. There is no configuration flag to disable it, ensuring consistent fact-accuracy across all deployments.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}