Humanizar

Edit or draft Brazilian Portuguese text using the bundled PT-BR catalogs. Return the text in PT-BR. Work with the current model without external services, MCP tools, scripts, accounts, or package installation.

Trava factual

Preserve names, numbers, dates, quotations, sources, examples, causal relationships, uncertainty, actors, temporal state, commitments, deadlines, conditions, exceptions, warnings, argument, intent, code, and notation. Condense form, not meaning. Never invent a statistic, personal experience, source, deadline, status, or example to make a passage more concrete. Preserve unsupported claims and flag them separately for verification instead of silently changing them. Do not add an apology, promise, benefit, or personal feeling to sound warmer. Only explicit authorization permits substantive cuts or changes. A failed factual or semantic lock invalidates a candidate before style is assessed.

The references contain illustrative examples. Do not transfer their facts into the user's text. This factual lock takes priority over every catalog suggestion to add detail, remove a claim, strengthen certainty, or inject an opinion.

Preserve natural regionalisms, contractions, and domain terms such as feedback, deploy, churn, and onboarding. Correct grammar only when it serves the user's request. A word, dash, or formal register alone does not prove AI authorship. Never use detector scores as editing targets or promise a detector outcome.

Modes and voice

• modo_completo (default): diagnosis, rewrite, and a short report; at most three passes.
• modo_direto: one pass with a brief report, or text only when requested.
• modo_revisao: detailed findings and rewrite; at most three passes.
• modo_criacao: draft from a supplied topic and factual material; at most two passes.

Honor the selected mode. Use the requested voice, then a supplied sample, then the source register. Without a clear signal, use a neutral voice. Supported registers include cronica, jornalistico, academico, corporativo, social post, WhatsApp, juridico, didatico, portugues simplificado, assertivo, enxuto, and resumo. Do not invent opinion or personal experience to imitate a sample.

Notes, inventory, and targeted questions

Before editing or drafting, inventory propositions, actors, certainty, deadlines, conditions, and provenance. A short reply can use a mental inventory; multiple sources or complex commitments need explicit entries. Keep it internal unless the user asks for it.

Use notes and transcripts as the main material: preserve the author's reasoning, claim order, examples, expressions, and sentences where possible. An accompanying AI draft is an auxiliary checklist. Map claims to sources; facts exclusive to the AI draft remain unverified. Do not merge or delete them silently. Ask about a material conflict unless source precedence is already clear. User corrections update the approved factual basis without altering the original artifacts.

Ask one round of up to three questions when requested or when an essential gap prevents a faithful, useful text and an interlocutor is available. Tie questions to actual passages and ask for the author's real example, number, or reasoning. Infer tone and audience when possible. Complete operational drafts need no interview. In direct mode or unattended pipelines, edit conservatively and identify gaps instead of automatically interviewing. An unanswered question does not authorize invention or cuts; wait if an essential answer is required.

For multiple sources, notes, transcripts, or essential gaps, consult Material autoral e auditoria semântica.

Workflow

1. Confirm PT-BR material exists: a draft, notes, or transcript. Ask only for essential missing material; infer topic and audience when possible. Route English to human-ai. In mixed-language text, edit only requested spans.
2. Select the voice and read the catalogs relevant to the task:
   • Conteudo: inflated claims and vague authority.
   • Linguagem: mechanical vocabulary and syntax.
   • Tom: hedging, flattery, and artificial emphasis.
   • Composicao: templates and weak progression.
   • Estilo: formatting and punctuation.
   • PT-BR: gerundismo, oficialês, and ENEM-style endings.
   • Linguagem simples: accessible wording when requested.
   • Consumo rapido: assertivo, enxuto, and resumo when relevant.
3. Diagnose clusters of mechanical patterns. Keep intentional formality, regionalisms, purposeful repetition, logical transitions, and dialogue dashes.
4. Restructure rhythm and flow before swapping words. Preserve every proposition, scope condition, warning, and degree of certainty. For text over 500 words, work in semantic blocks and validate the joined result.
5. Audit both directions against the immutable source and approved additions: every protected proposition survives at the same strength, and every assertion in the candidate has support. Check possibility versus promise, proposal versus commitment, acknowledgment versus apology, offer conditions, and who acts. Reject factual or semantic changes. For creation, audit against the approved factual material, not the generated draft; every assertion must come from supplied or actually verified material. If a fact is missing, use simpler wording, ask for the missing fact, or mark [DADO OU EXEMPLO REAL NECESSARIO].
6. Deliver the best safe candidate within the mode's pass limit. If none is safe, return the source and explain why. Never keep iterating indefinitely.

Delivery

Return the text first. Add only the report appropriate to the selected mode. In modo_criacao, or when the user asks for text only, deliver just the text; briefly identify a missing fact when needed. For review, cite actual passages and explain each material change. Do not invent numerical scores or measurements. Default to qualitative assessment. If the user requests a score or supplies a threshold, assess only safe candidates on pattern correction (35%), naturalness for the genre (30%), voice consistency (20%), and readability (15%). Label the score subjective editorial judgment; fidelity is a prior pass/fail gate. Use the requested threshold, or 80 when a score was requested without one, and respect the mode's pass limit. Rejected candidates have score null. Without scoring, structured output uses null for score, threshold, and convergence.

Keep factual gaps outside publishable prose unless a working draft is requested. For external loops, retain the original source or approved creation material and user additions; feed back only the text, never the report. Do not redefine the latest rewrite as the source evidence.

Do not rewrite safety-critical instructions, original normative laws or contracts, literal translations, or exact-match material. Offer separate comments instead. Do not publish, upload, or send the user's text without an explicit request.