humanizer-stack — скилл

---
name: structural-humanizer
description: >-
  Remove the discourse-level (structural) signs of AI writing that survive surface
  editing: stated lessons and moral-of-the-story closers, tidy single-track arcs,
  embodied-emotion performance ("chest tightened"), vague allusions instead of named
  references, unbroken linear structure, and shape convergence across pieces. Grounded
  in the StoryScope study (Russell et al. 2026): narrative structure alone detects AI
  text at 93.2% F1, and professional stylistic rewriting moved detection only 1.6
  points. Use as the SECOND pass after the humanizer skill (which handles words and
  phrasing) whenever writing or revising LinkedIn posts, course lessons, blog posts,
  essays, newsletters, or emails that must read as human. Triggers: "humanize",
  "de-slop", "AI tells", "make this sound human", "structural pass", "deep humanize".
---

# structural-humanizer

Read this first. The `humanizer` skill fixes words: "delve", em dashes, rule of three,
negative parallelism. This skill fixes what survives that pass: the structure. The two
are different jobs, run in sequence. Surface pass first, structural pass second.

**Why this layer matters more.** StoryScope (Russell et al. 2026, arXiv:2604.03136)
classified 61,608 stories from humans and 5 LLMs using only discourse-level features,
with all style features withheld: 93.2% detection accuracy. Then they ran AI text
through LAMP, a professional span-level rewriting framework that removes cliche,
purple prose, and redundant exposition (functionally, a surface humanizer). Detection
dropped 1.6 points. Meanwhile the surface layer is decaying on its own: GPT 5.4
already slashed em-dash usage, and fine-tuning drops stylistic detection from 97% to
3%. The durable fingerprint is structural, and fixing it requires structural rewrites,
not word swaps. Full findings with numbers: [references/storyscope-findings.md](references/storyscope-findings.md).

## The trap (same trap as unslop-ui)

Do not replace one default with another. If every piece now opens mid-scene, names
three feelings, and ends unresolved, that is a new detectable cluster. The study's
deepest finding is convergence: all five AI models occupy one tight region of
structural space while humans are dispersed and rare. Rarity IS the human signal.

So: **pick 1-2 structural interventions per piece, vary them across pieces, and be
able to say why this piece got this shape.** Never apply the whole menu at once.

## The six audits