Every guide to de-AI-ing a draft hands you a list of words to delete. That list starts expiring the day it ships: naming a tell is what kills it. Underneath it sits the failure no list touches — the draft commits to nothing. So I built an editing skill that fixes that one first, and put it on GitHub.
Before
In today's fast-paced development landscape, it's worth noting that observability has become paramount. Our latest release introduces a robust set of enhancements to the query planner, which now seeks by trace ID rather than scanning every span in the window, reducing median trace lookup from 4.2s to 380ms and empowering teams to delve into their traces with unprecedented clarity.
After
Median trace lookup went from 4.2s to 380ms. The old query planner scanned every span in the window; the new one seeks by trace ID first. If you page through traces during an incident, you will feel it.
Every number and every mechanism in the second version is in the first. The edit cut what was standing in front of them.
The list dies because someone published it
When delve, intricate and showcasing were publicly named in early 2024, their frequency in arXiv abstracts began falling around March — before GPT-4o shipped that May. Writers killed those words, not a model update. Meanwhile significant and additionally kept climbing, because nobody had named them (Geng and Trotta, ~1.29M abstracts).
So a list you can read today is already decaying, and nobody can tell you how fast: the diffusion timelines that circulate as fact have no source behind them.
What doesn't expire
Across seven instruction-tuned models and five registers, every model's distribution of 67 lexico-grammatical features sits outside the human-to-human confidence band. Reword the prompt and the absolute distance moves; the ranking does not, at correlations of at least 0.985 (Nieth et al.).
"Write more casually" does not reach the thing that is broken.
The fourth is the only one editing cannot repair, and it is getting worse. Fabricated references ran 1 paper in 2,828 in 2023, 1 in 458 in 2025, and 1 in 277 in the first seven weeks of 2026 (Topaz et al., Lancet 407(10541):1779–1781). Retrieval changed the failure's shape rather than fixing it. The modern version cites a real, resolvable paper that does not support the sentence it is attached to.
The skill
bluepencil edits a draft, or audits one without touching a word. It writes a voice contract first: the core point, plus three to five phrases quoted verbatim from your draft that it has promised to protect. Then it edits, then it runs a 36-item rubric against its own output. On long or published work it escalates to two critics in fresh context, one hunting voice loss and one hunting surviving slop. Voice-loss findings win ties.
That last stage exists because self-scoring is measurably biased, and not even in a consistent direction: LLM judges deviate from +9.40% to −16.64% on their own output depending on the model (Ye et al., CALM, ICLR 2025). You cannot correct for that by assuming leniency. You correct for it with a reader who holds the artifacts and none of your reasoning.
Two things it refuses to do. It will not optimise against an AI detector: a reinforcement-learned paraphrase attack drops detector true-positive rate to 0.024 at 1% false positives (arXiv:2602.08934), and seven detectors averaged 61.3% false positives on TOEFL essays written by non-native speakers (Liang et al., Patterns 4(7):100779). Writing toward a classifier makes the prose worse and lands hardest on people who did nothing wrong. And it will not guess who wrote something. It names patterns you can check yourself.
Install
One line, any agent:
npx skills add chernistry/bluepencilClaude Code, as a plugin:
/plugin marketplace add chernistry/bluepencil/plugin install bluepencil@bluepencilOne line, no tooling:
git clone -q --depth 1 https://github.com/chernistry/bluepencil /tmp/bp && mkdir -p ~/.claude/skills && cp -r /tmp/bp/skills/bluepencil ~/.claude/skills/ && rm -rf /tmp/bp && echo "bluepencil installed"Then: Blue-pencil this post. Or: Audit this draft — don't rewrite it.
The thing that convinced me the fourth tell is the real one
The evidence file behind the skill was assembled from six independent deep-research runs answering one prompt. Five of them reported a specific set of questions as unmeasured — how many critic rounds before gains stop, whether a written voice contract beats an unconstrained edit, what human sentence-length variance actually is.
The sixth answered every one of those questions with a precise figure and an arXiv identifier. Voice preservation of 88.7%. Fabricated criticisms reaching 42% by round four. None of those papers exist.
The most useful-sounding numbers in the entire corpus had been generated to fill exactly the holes where no measurement was available. In a document explicitly instructed to open every source. That is why the reference file ships as three separate lists: what was measured, what is folklore wearing a number, and what nobody has measured.