What "Weak Alone" Means in Humanizer Pattern Definitions

In the Humanizer project, "weak alone" indicates that a single instance of an AI-generated tell is insufficient to trigger an edit and must appear alongside other tells in the same passage before the system will act on it.

The open-source blader/humanizer repository provides a framework for identifying AI-generated text through numbered pattern recognition. Within the project's core definitions, the term "weak alone" categorizes specific linguistic cues that lack standalone diagnostic strength but collectively signal synthetic prose. According to SKILL.md, these marginal cues require companionship from other tells before warranting editorial action.

Understanding Pattern Strength and the "Weak Alone" Flag

Humanizer organizes AI-generated tells into numbered patterns ranging from the strongest (§1) to the weakest (§5). The "weak alone" designation serves as a critical safeguard against over-editing clean prose based on isolated, low-impact cues.

According to the source code in SKILL.md at line 29, the skill's introductory rules state that patterns marked weak alone "needs company from other tells in the same passage before you act." This requirement ensures that single instances of marginal punctuation or phrasing—such as an em-dash or a hyphenated buzzword—do not trigger unnecessary rewrites in otherwise human-authored text.

Common "Weak Alone" Pattern Examples

Throughout SKILL.md (lines 62-63 and 366-367), the documentation repeatedly emphasizes that marginal cues carry the weak alone tag. These include:

• Dashes (em-dashes and en-dashes)
• Curly quotes (smart quotation marks)
• Stacked qualifiers (phrases like "to be fair," "might arguably," or "could potentially")
• Hyphenated pairs (buzzword constructions like "cross-functional," "data-driven," or "high-quality")
• Passive voice constructions

The documentation explicitly states: "Act on a weak alone tell only when several tells share a passage."

Implementing "Weak Alone" Detection Logic

While Humanizer itself is implemented as a Markdown skill, the detection logic can be replicated in any programming language. The key principle requires identifying at least two distinct weak-alone patterns before triggering an edit, embodying the "needs company" rule from the source documentation.

The following Python implementation demonstrates this collective detection approach:

# Example: Detecting whether a tell should be acted upon

WEAK_ALONE_PATTERNS = {
    "dash": r"—|–|--",
    "curly_quotes": r"[“”]",
    "stacked_qualifiers": r"\b(?:to be fair|might arguably|could potentially)\b",
    "hyphenated_pairs": r"\b(?:cross-functional|data-driven|high-quality)\b",
    "passive_voice": r"\b(?:is|are|was|were) \w+ed\b",
}

def find_tells(text):
    """Return a list of (pattern_name, match_span) for all tells found."""
    tells = []
    for name, regex in WEAK_ALONE_PATTERNS.items():
        for m in re.finditer(regex, text):
            tells.append((name, m.span()))
    return tells

def should_edit(passage):
    """Only edit if at least two different weak‑alone tells appear."""
    tells = find_tells(passage)
    unique_patterns = {name for name, _ in tells}
    return len(unique_patterns) >= 2   # need “company”

# Usage

paragraph = "The new policy — announced without warning — affects thousands."
if should_edit(paragraph):
    # Perform rewrite (e.g., replace dash with commas)

    rewrite = re.sub(r"—", ",", paragraph)
    print(rewrite)

The should_edit function embodies the weak alone principle by requiring two or more distinct weak-alone tells before any modification is performed, preventing false positives from single, isolated cues.

Summary

• "Weak alone" marks AI-generated tells in Humanizer that require multiple instances before triggering edits.
• Patterns range from §1 (strongest) to §5 (weakest), with weak-alone cues falling in the marginal categories.
• The rule requires "company from other tells in the same passage" as defined in SKILL.md at line 29.
• Common examples include dashes, curly quotes, stacked qualifiers, hyphenated pairs, and passive voice.
• As documented in SKILL.md (lines 62-63, 366-367), only passages containing several weak-alone tells warrant rewriting.

Frequently Asked Questions

What does "weak alone" mean in Humanizer?

In Humanizer, "weak alone" describes an AI-generated tell that is not strong enough individually to justify an edit. According to SKILL.md, these patterns need to appear alongside other tells in the same passage before Humanizer will act on them, functioning as a safeguard against over-editing.

Which patterns are marked as "weak alone" in the source code?

The SKILL.md file marks several marginal cues as weak alone, including em-dashes and en-dashes, curly quotation marks, stacked qualifiers like "to be fair" or "could potentially," hyphenated buzzwords such as "data-driven," and passive voice constructions. These are considered low-impact individually but significant collectively.

How many "weak alone" tells are required to trigger an edit?

While the exact threshold depends on implementation, the source logic in SKILL.md and the reference Python implementation demonstrate that at least two different weak-alone tells must appear in the same passage. The documentation emphasizes requiring "several tells" to share a passage before acting on any of them.

Where is the "weak alone" rule defined in the Humanizer repository?

The primary definition resides in SKILL.md at line 29, which states that patterns marked weak alone "needs company from other tells in the same passage before you act." Additional clarifications appear at lines 62-63 and 366-367, reminding users to "Act on a weak alone tell only when several tells share a passage."

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