humanizer-pl - Polish and English AI Text Humanizer for Codex, GPT and ChatGPT

AI writes smooth. People write specific.

humanizer-pl is a bilingual AI text humanizer for Codex, GPT and ChatGPT. It removes common AI-writing tells from Polish and English while preserving meaning, facts, citations, terminology and the author's voice.

Polish is not translated English. GPT and Claude have their own Polish tells: calques, adjective chains, too many participles, inflated openings and choppy parataxis. This skill checks both languages and changes the text only when there is a reason to do so.

TL;DR (EN): A bilingual de-slop skill for Codex, GPT and ChatGPT. It auto-detects Polish or English, removes AI-writing tells, preserves meaning and voice, and runs layered output QA. The English patterns come 1:1 from blader/humanizer; the Polish layers, Bielik advisor and eval harness come from the upstream Polish project.

Install · See it · Use it · What it checks · Bielik · Attribution

See it

AI version:

Dzień dobry, w nawiązaniu do Państwa zapytania, pragnę uprzejmie poinformować, że nasz dedykowany zespół dołoży wszelkich starań, aby kompleksowo sprostać Państwa oczekiwaniom.

Human version:

Dzień dobry, sprawdziłem Państwa zapytanie - zrobimy to do piątku. Cena bez zmian, 4 200 zł netto. Jakby cokolwiek było niejasne, proszę pisać. Pozdrawiam.

The full bank contains 15 Polish pairs plus the original English pairs from blader/humanizer: references/przyklady.md.

Install

Codex marketplace - recommended

Add this marketplace in Codex:

https://github.com/mateuszkozlowski/codex-humanizer-pl

The marketplace marks Humanizer PL for automatic installation. If you added the marketplace but the plugin is not active yet, run:

codex plugin add humanizer-pl@codex-humanizer-pl

After installing a plugin, open a new Codex task and invoke $humanizer-pl.

Standalone skill

git clone https://github.com/mateuszkozlowski/codex-humanizer-pl \
  ~/.codex/skills/humanizer-pl

The default mode has no dependencies. The optional Bielik mode is local and requires Ollama plus a model download.

Use it

In Codex, use the full name:

$humanizer-pl:humanizer-pl

Or the short name:

$humanizer-pl

Examples:

zhumanizuj ten post, zachowaj fakty i mój ton
wywal AI-slop z tego maila
przepisz ten tekst naturalniej, ale nie zmieniaj terminologii
zhumanizuj + sprawdź Bielikiem
zhumanizuj + raport HTML
zapamiętaj mój żargon

The skill detects PL/EN automatically. It does not require a separate command for each pattern.

What it checks

One pass is not enough. The skill routes the text through separate checks and stops before it silently changes something that matters.

Layer	What it does
Language	Detects Polish or English and loads only the relevant pattern set.
De-slop	Removes inflated openings, promotional filler, vague attributions, chatbot politeness, rule-of-three phrasing, negative parallelisms and other recurring tells.
Polish naturalness	Checks calques, lazy anglicisms, adjective chains, participles and parataxis; preserves deliberate professional jargon.
Structure	Rebuilds sentence rhythm where needed instead of only swapping punctuation or synonyms.
Protected content	Keeps numbers, URLs, citations, placeholders, names, terminology and the requested language.
Final gates	Checks Polish diacritics, forbidden dash characters, non-empty output and factual identity.

For longer texts, an additional selective catalogue covers meta-commentary, artificial rhetoric, duplicated conclusions, enumeration and formatting: references/advanced-tropes.md.

What it does not claim

This is a writing aid, not an authorship detector. One word, one dash or one sentence structure is not proof that a text was generated by AI. Unusual human writing should survive the pass. The goal is a more natural draft, not a verdict about who wrote it.

Optional Bielik reviewer

The default reviewer is the main model, so the skill works everywhere with no setup.

In the optional Bielik mode, a local Bielik model through Ollama can flag possible Polish calques or suggest alternatives. Bielik does not rewrite the text automatically: the main model evaluates each suggestion, shows risky changes and leaves the final decision to the user.

Install only when you explicitly want this mode:

./scripts/install-bielik.sh

The installer explains that Ollama and the model may download several gigabytes. Setup details: references/setup-bielik.md.

Personal jargon profile

The skill does not hard-code a legal, marketing or technical dictionary. You can ask it to zapamiętaj mój żargon or zbuduj mój profil żargonu.

After confirmation, the private list is stored locally in references/jargon-profile.local.md and is ignored by Git. The profile helps preserve terms such as cesja, rękojmia, lead or funnel instead of translating or smoothing them away.

Checks and evals

Run the deterministic checks from the repository root:

python3 evals/run_evals.py
python3 evals/run_qa.py

run_evals.py preserves the original binary assertions. run_qa.py adds regression cases and checks numbers, URLs, language, placeholders, protected fragments and empty output. These are output-quality gates, not a classifier of human versus AI authorship.

What's in the box

humanizer-pl/
  .agents/plugins/marketplace.json       Codex marketplace manifest
  plugins/humanizer-pl/                  Codex plugin package
    .codex-plugin/plugin.json
    skills/humanizer-pl/                 materialized standalone skill
  SKILL.md                               language router and process
  references/
    patterns-en.md                       English patterns from blader/humanizer
    patterns-pl.md                       Polish pattern catalogue
    polszczyzna-pl.md                    Polish naturalness layer
    advanced-tropes.md                   extended selective review
    przyklady.md                         original before/after example bank
    raport-html.md                       optional side-by-side report format
    qa-checklist.md                      manual review checklist
  evals/                                  regression and QA runners
  agents/openai.yaml                     Codex skill metadata
  scripts/                               optional local Bielik tools

Privacy and boundaries

The skill has no account, hosted service or telemetry of its own. The default mode only provides instructions to the model running the Codex task. The optional Bielik reviewer talks to Ollama on your machine by default (localhost:11434).

Do not treat any humanizer as a guarantee that a text will evade an AI detector. Review the result, especially when facts, quotations, legal language or academic claims matter.

Attribution

This Codex fork is based on pielas-activy/humanizer-pl (MIT). The original project supplies the Polish layers, Bielik/Ollama integration and evals. Its English patterns come from blader/humanizer, which is based on Wikipedia's Signs of AI writing guide.

The additional trope catalogue is curatorially informed by tropes.md, and the blind-calibration procedure by AI or not quiz. This Codex port adds agents/openai.yaml, materialized plugin paths, layered QA and safer opt-in handling for local tools. None of these sources is treated as evidence of authorship. See LICENSE for the complete licensing and attribution terms.

License

MIT. See LICENSE.