p3n

The AI pen: writing skills for your AI agent.

Skill	What it does
AI Detection	Was text written by a person or a model? A score, a label (human, AI, AI-edited, humanized, mixed), a verdict, and the sentences and patterns that gave it away.
Deslop	Makes text say what it means: cuts filler, restatements and benefit tails, and asks for the facts a reader is missing.
Editorial	Deslop, but with editorial context (e.g., technical, finance, etc.) and less AI sounding result.
Humanize	Rewrites text to register less AI-generated, without changing what it says.

Install

Each folder in skills/ is an agent skill: a SKILL.md, the rules and prompts, plus small scripts. For Claude Code, link the skills into ~/.claude/skills/:

git clone https://github.com/zdresearch/p3n && cd p3n
for s in skills/*; do ln -s "$PWD/$s" ~/.claude/skills/; done

Then ask your agent to "check draft.md for AI writing" or "run editorial on draft.md".

The scripts need Python 3.11+ with no dependencies. AI Detection and Humanize also use tropa-mini, an open-weights classifier, as AI detection oracle, plus PyTorch and about 2 GB of disk and memory:

pip install -r requirements-model.txt

How it works

Your AI agent does the reading and writing (as usual), and hands the editorial, rewriting and checking parts to subagents on a small fast cheap model (Claude Haiku 5.5, or GPT-6 Luna at low reasoning effort). A subagent has fresh context and reads the text cold, avoiding context creep and sloppiness. It is also fast, cheap and efficient enough for reading/writing.

AI Detection runs three checks, then a judge. The first check is a linter with ~60 rules that model-written text shows far more often than human text (rules.toml, each with a score). The second check uses the LLM rules from the same file via a subagent, for rules that are not static. AI detection oracle classifies the text as the third step. Then the judge, another subagent, reads the text and the result of the three checks and issues the verdict score, the label, the verdict and the flagged sentences.

Deslop and Editorial split the text one sentence per line and label every line: P3S2 is sentence 2 of paragraph 3. An editor subagent reads the labelled file with an audience brief in mind (i.e., who reads this text, and what for) and returns, per label, a change, a cut, a question or a note, plus a score of 0 to 100. The writer agent (your agent that is running the skill) answers the questions, applies the changes as seen fit, and sends the reworked text back. The writer and the editor iterate in rounds until the score reaches 80 or 5 rounds. The editor doesn't modify the text itself, it just asks questions and makes suggestions like a real-world editor. The editorial rubrics are simple markdown (deslop, editorial), they can be modified and tuned to specific work.

Humanize rewrites each paragraph by ~12 rules (rules.md). With AI detection oracle installed, it writes three versions of each paragraph with different emphasis, checks them, and keeps the lowest AI version.

Hosted API

All skills are available as an API at https://api.p3n.si/v1 and explained in more detail at p3n.si. Multiple free calls per hour, with a paid model that is still very affordable:

curl https://api.p3n.si/v1/detect -H 'Content-Type: application/json' -d '{"text": "Paste AI generated text here."}'

The skills can use the API as well (provide them with an API key). This route is especially useful for AI detection and Humanize skills that require setting up oracles locally and may be slow. We use the online versions ourselves despite having them installed locally! The local versions are only worth the effort if you are going to use the skill with significant volume.

Versions

skills/<skill>/VERSION names the rules and prompts this repository holds. The hosted API runs the same ones under the same name, with the models they were tuned on; pin it with "options": {"version": "detect-v1.4"}.

AI Detection Oracle

AI Detection and Humanize use tropa-mini by wasitaigenerated.com as the AI detection oracle: a DeBERTa-v3-large classifier with four classes (human, AI, AI-edited, humanized), Apache-2.0, derived from desklib/ai-text-detector-v1.01 and microsoft/deberta-v3-large (both MIT). It is downloaded from Hugging Face on first use.

Tests

python3 -m unittest discover tests

License

MIT.