revisao-humanizada

name	revisao-humanizada
description	Humanizes narrative copy by removing AI tells, varying rhythm, and stripping templated phrasing before publish
version	0.1.0

Revisao Humanizada

Final pass on any narrative copy headed for a human audience. Strips the typical AI fingerprints: em-dashes, triadic patterns, hedging, monotone sentence length, and recycled connectors. Mirrors the global humanizer skill behavior, scoped to the marketing-engine pipeline.

When to invoke

• Right before publishing any caption, post body, blog article, email, ad copy, video script, landing page section.
• After the first draft from copywriter-curto, caption-multi-platform, or any LLM-generated long-form text.
• When a reviewer flags copy as sounding generic, robotic, or template-like.
• During QA pass on a scheduled piece in .specs/pieces/.
• Whenever the brand voice score from a manual read drops below the threshold defined in .specs/strategy/BRAND.md.

Inputs

• text: string. The draft copy to humanize.
• language: string. Target language code (pt-BR, en, es).
• platform: string, optional. Platform context (ig, tiktok, youtube, linkedin, email, blog).
• brand_voice_path: string, optional. Path to .specs/strategy/BRAND.md for tone constraints.
• preserve_terms: array of strings, optional. Brand terms, product names, and CTAs that must stay verbatim.

Process

1. Read text and inventory AI tells: em-dashes, sentences starting with In conclusion, Moreover, Furthermore, triads like fast, simple, and effective.
2. Read brand_voice_path if provided. Note the brand tone, banned words, and required phrasing.
3. Call llm-router with task: humanization. Send the draft, language, platform, brand voice excerpt, and the preserve_terms list.
4. Apply rewrites: vary sentence length (short, medium, long, fragment), break triadic patterns, replace em-dashes with comma or period, remove hedge words (perhaps, it could be argued).
5. Verify preserve_terms still appear verbatim in the output.
6. Confirm the output language matches language.
7. Score the result against an internal AI-tell checklist. If three or more tells remain, run one more pass.
8. Return the humanized text plus a short diff of what changed.

Outputs

• text: string. Humanized copy.
• changes: array of strings. Bullet log of edits applied.
• ai_tells_remaining: integer. Count of tells the final pass could not remove.
• passes_used: integer. How many humanization rounds ran.

Examples

Example 1: caption rewrite

Input: { text: "In conclusion, color analysis is fast, simple, and effective.", language: "en", platform: "ig" } Output: { text: "Color analysis works. It is quick. The results stick.", changes: ["removed In conclusion", "broke triad fast/simple/effective"], ai_tells_remaining: 0, passes_used: 1 }

Example 2: long-form blog paragraph

Input: a 200-word draft full of em-dashes and triads. Output: same paragraph rewritten with mixed sentence rhythm, em-dashes converted, brand terms intact.

Failure modes

• Preserve_terms vanish in rewrite: re-inject them in the exact original positions and rerun humanization with stronger constraint.
• Output language drifts: force the language constraint in the next call and reject the previous result.
• Result still scores high on AI tells after two passes: surface the issue, return the best result, and flag for human edit.

Related skills

• copywriter-curto: produces the draft this skill polishes.
• caption-multi-platform: each variant is humanized before being scheduled.
• llm-router: actually runs the LLM call for the humanization pass.
• compliance-generic: should run after humanization to confirm no banned phrasing slipped in.