harness-humanizer

name

harness-humanizer

description

Use when prose reads like AI — to remove "AI slop" (empty hedging, listicle stems, smooth transitions that hide the absence of a claim, generic filler) and rewrite it into writing with a real point of view. Trigger on requests like "humanize this", "de-slop", "remove the AI slop", "make this sound less like AI / less like ChatGPT", "this reads like AI", or after generating prose that needs a quality pass. Detects, rewrites the fixable parts, self-scores against an embedded rubric, and iterates to a bar — preserving meaning exactly, flagging hollow spans instead of inventing claims, and reporting changes rather than overwriting.

Harness Humanizer

Turn AI-slop prose into writing that survives a hostile editor's red pen — without swapping one kind of slop for another.

Two hard rules (read first)

1. Fidelity over flair. Preserve the original meaning and claims exactly. Only subtract hedging/filler and sharpen what is already there. Never inject stance, edginess, em-dash theatrics, or first-person personality the content did not earn. Swapping AI-slop for edgy-slop is a failure, not a fix.
2. Flag hollow spans, don't fabricate. Some prose is weak because it has no point to make — rewording cannot save it. Flag those. Do not invent a hot take to make them sound sharp.

The loop

0. Scope. Work paragraph by paragraph. Skip code blocks, blockquotes, headings, and genuine lists.

1. Pre-flag. Run the cheap deterministic pass to narrow attention:

python3 scripts/flag_slop.py <file>     # or: cat text | python3 scripts/flag_slop.py

It returns JSON spans (hedge stems, listicle openers, em-dash density, "in today's…", filler intensifiers, etc.). These are candidates, not verdicts — you still judge every paragraph.

Optionally, python3 scripts/flag_slop.py --score <file> returns a per-paragraph slop_band. Treat it as a surface-tell meter, not a humanness score: it measures how many slop patterns appear, not whether a real claim is present. A paragraph with zero tells can still be hollow and fail the rubric — so a high slop_band never excuses you from step 2. See references/slop-catalogue.md for which tells the detector can and cannot see.

2. Judge. Score each paragraph against references/rubric.md → strong | moderate | weak | fail, with a one-line reason. The bar is the hostile-editor test: would this survive a red pen? does removing it lose anything?

3. Triage each paragraph below strong:

• Rewordable — there's a real claim buried under hedging/filler → rewrite.
• Hollow — weak because there's no actual point → flag, don't fabricate.

4. Rewrite the rewordable ones, applying references/guardrails.md. Subtract the hedging, sharpen the existing claim, keep the meaning identical.

5. Self-score the rewrite against the rubric again.

• Reached strong → lock it in.
• Still below → iterate (back to step 4). Maximum 3 passes total.
• After 3 passes still not strong → keep the best version and flag it ("couldn't reach strong — may need a real claim, not better words").

6. Report — do not overwrite. Return three things:

• Humanized text — rewrites applied; hollow spans left intact.
• Change log — per paragraph: before-band → after-band and what changed.
• Flags — hollow spans + any span that hit the 3-pass cap.

The human or calling agent decides what to accept.

Properties this loop must preserve

• Fail-honest: hollow and capped spans are always surfaced, never quietly "polished."
• Idempotent: prose that already scores strong is returned unchanged.
• Non-destructive: you produce a report + change log, not an in-place edit.

References

• references/rubric.md — the scoring bands, slop indicators, the two tests, and the rewordable-vs-hollow triage rule. Load it for step 2.
• references/guardrails.md — fidelity rules and the over-correction anti-pattern catalogue. Apply it for step 4.
• references/examples.md — before→after pairs, "flag don't fabricate" cases, and over-correction PASS/FAIL pairs. Consult when a rewrite is non-obvious.
• references/slop-catalogue.md — the full taxonomy: every tell, why it reads as AI, the detector type that catches it (or why none can). The map of the detector's blind spots.