Does Humanizer Change Facts or Add New Ones? Inside the blader/humanizer Accuracy Guarantee

Humanizer is architected to rewrite AI‑styled prose while strictly preserving every factual claim, explicitly forbidding the invention of new details and prompting users for clarification when source information is missing.

The blader/humanizer repository provides a Markdown skill that transforms robotic, AI‑generated text into natural human voice without altering underlying facts. Unlike generative models that might hallucinate dates, names, or statistics, Humanizer operates under a zero‑fabrication constraint hardcoded into its prompt logic. Understanding how this tool handles factual integrity requires examining the specific source files that enforce its "keep what it says" mandate.

Core Factual Preservation Rules

The skill's foundational safety guardrails are defined in two key documentation files that explicitly prohibit factual modification.

According to SKILL.md at line 15, the core instruction states: "Keep what it says. Do not make anything up." This directive is not merely suggestive—it functions as the primary constraint governing all rewrite operations. The prompt architecture treats any deviation from source facts as a critical error.

The README.md reinforces this policy at lines 73‑74, clarifying that "Humanizer … does not make things up." The documentation specifies that any name, number, date, quote, citation, or factual detail must come from the source; if a sentence requires a detail that is absent from the input, Humanizer asks instead of inventing one.

The Four-Step Verification Workflow

Humanizer implements a strict processing pipeline designed to catch and block any factual drift before output generation.

Mark the Tells

In the initial analysis phase, the skill identifies AI‑style patterns—such as overly formal transitions or repetitive structures—while leaving every factual claim completely untouched. This step strictly separates stylistic markers from content accuracy.

Draft the Rewrite

During rephrasing, the draft logic contains embedded safeguards that block any addition of unsupported facts. The skill restructures sentences and varies vocabulary while retaining all supported data points exactly as they appear in the source material.

Check the Draft

A validation layer scans the drafted rewrite for added or omitted factual elements, flagging any deviation as an error. This verification ensures that no dates have been shifted, no names have been changed, and no statistics have been altered during the stylistic transformation.

Write the Final Version

The final output emits only after the check passes, guaranteeing that no new facts have been introduced and no existing facts have been modified. Because Humanizer is purely a Markdown skill, all processing occurs locally within the prompt logic without runtime dependencies or external APIs that could alter content.

Usage Examples

Humanizer integrates with chat‑compatible agents and command‑line interfaces while maintaining its factual constraints.

Calling Humanizer directly from a chat agent:

/humanizer

Please humanize this text:
[Paste AI‑generated prose here]

Humanizing a file via CLI after installing the skill globally:

npx skills add blader/humanizer --global   # install the skill

humanizer docs/launch-post.md               # rewrite only the prose in the file

Providing a writing sample to match voice:

/humanizer

Here's a sample of my writing for voice matching:
[Paste 2‑3 paragraphs of your own writing]

Now humanize this text:
[Paste AI‑generated prose]

Handling Missing Information

When source material lacks required details, Humanizer does not fabricate plausible data. Instead, it surfaces the gap to the user for resolution.

For example, if input text references an unspecified date:

/humanizer

[AI text that mentions an unspecified date]

Humanizer responds with a clarification prompt such as: "The source does not specify the date—please provide it or omit the reference." This behavior is enforced by the validation logic in scripts/validate-package.py, which checks that the factual‑preservation policy remains present in all skill distributions.

Summary

• Humanizer never changes facts or adds new ones; it operates under a strict "keep what it says" rule defined in SKILL.md.
• The four‑step workflow (Mark, Draft, Check, Write) includes validation layers that explicitly block factual drift.
• Missing information triggers user prompts rather than hallucinations, ensuring no fabrication occurs.
• As a local Markdown skill without external API dependencies, all processing stays contained within the audited prompt files (agents/openai.yaml, SKILL.md, README.md).

Frequently Asked Questions

Can Humanizer correct factual errors in the source text?

No. Humanizer treats the input as the authoritative source of truth and will not correct perceived inaccuracies. If the source contains an incorrect date or statistic, that error will persist in the humanized output. The tool's sole function is stylistic transformation, not fact‑checking.

What happens if the AI text contains placeholders or incomplete data?

Humanizer detects missing factual details during the "Check the Draft" phase and halts processing. Rather than inventing plausible fill‑ins, it returns a request for clarification. Users must either supply the missing information or remove the reference before the tool generates final output.

Does Humanizer access external APIs to verify facts?

No. According to the repository architecture, Humanizer is a static Markdown skill that processes text locally. The agents/openai.yaml file merely exposes the skill to compatible agents; there are no runtime dependencies or external verification services that could modify content during execution.

How can I verify that Humanizer hasn't altered my content?

You can audit the scripts/validate-package.py validator, which checks that the factual‑preservation policy remains intact in the skill files. Additionally, comparing the original and humanized outputs side‑by‑side will show that all names, dates, numbers, and quotes remain identical, with only sentence structure and word choice changing.

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