Skip to content

Proposal: a "light-edit" mode for human-written input (Mode A / B split) #172

Description

@vincent-wen789

The gap

The skill assumes its input is AI-generated — the whole prompt is written for "rewrite freely, inject voice, cover every paragraph." That's exactly right when someone hands you slop.

But a large share of real usage is people running humanizer on their own writing to "clean it up" or "make it read better." On human-written input, rewrite-everything is the wrong reflex: it sands the author off — polishing a specific, quirky voice into a smooth, sourceless one. The skill has no guardrail for this today; it treats a person's prose the same as a machine's.

Proposal: name two modes explicitly

Decide the job before touching anything, because the two modes pull in opposite directions:

  • Mode A — Humanize AI-generated text (current default). Danger = leftover AI tells. Rewrite freely. This is what the skill already does.
  • Mode B — Light-edit human-written text. Danger flips: the threat is over-editing. Rules for this mode:
    • Subtraction first — cut what genuinely trips the reader; don't swap a plain line for a "better" one. Clear obstacles, don't improve.
    • Hard cap — a passage needing more than ~3 changes is rare. If you're making ten, you've quit editing and started rewriting someone's voice.
    • No crafted phrases — a neat parallelism or a metaphor you added is the editor performing, not the author speaking. That's its own AI tell.
    • Author keeps veto over every change.

The governing test (useful in both modes, decisive in Mode B)

Human writing carries three things AI flattens — before any cut, check you're not erasing one:

  • Position — a specific person speaking from one time and place, not an interchangeable everyman view. "I saw X on the 7am train" beats a generic claim.
  • Cost — signs of real observation and effort behind the words; hard-to-fake specifics. AI generates frictionlessly and it shows.
  • Handwriting — the quirks, cadence, and repetitions that make a voice identifiable, including the "redundant" bits that carry feeling.

Then the one question that governs Mode B: after this cut, is the speaker still in the room? If the edit makes it sound like anyone — or a brand, or a press release — wrote it, revert it. "Cleaner" isn't the goal; "still them" is.

When unsure which mode applies, default to Mode B's restraint — over-editing a human is harder to undo than leaving one AI tell.


I've been running this as a local section for a while; it meaningfully changes behavior on human-written input. Posting as a discussion first since it changes the skill's default framing — happy to open a PR if you'd want it as a section. Feel free to rework the wording in your own voice.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions