Why a voice profile didn't get Claude past Substack's AI check

Question: can a model write a newsletter post that reads as its author's own work, if it is grounded in the author's research, steered by the author's opinions, and written against a voice profile built from the author's own writing? The test case: short Substack posts that summarise the research on this site and send readers back to it.

Short answer: no, not against Substack's detector. The first post, drafted by Claude from two research docs and three bullets of the author's opinions, passed every check built for it (a lint for AI vocabulary and phrasing, the voice profile, the author's own edits) and then scored AI 100%, Human 0% on Pangram, the classifier Substack runs on every post. That result is what Pangram's own technical report predicts. It is a learned classifier trained on AI rewrites of human documents, and on its authors' tests it catches humanizer rewrites 97.67% of the time and style imitation almost as reliably. A voice profile changes the surface of the text; the model's word-by-word choices underneath stay the same. The workflow now inverts: the model outlines, the author writes every sentence, and the model only reviews.

csarko.log is the Substack newsletter for this site. The plan was to turn one to four published research docs into a 300 to 600 word, plain-language post that links back to them, as a way to bring readers here. A Claude Code skill in this repository was built to do it, with four parts:

• The author's angle. Before drafting, the author gives about five bullets: why these docs, what he thinks, one concrete moment. The skill is barred from adding any claim, opinion or anecdote that isn't in the docs or the angle.
• A voice profile. A private file built from three Medium posts he wrote in 2019 and 2020 [1], about 26,000 words of his own messages to Claude Code (automated runs filtered out), and an eight-question interview. It records his habits: short sentences, colons for the reveal ("got my first grade: F"), dry humour aimed at himself, no hype words, emoji only at a punchline, and a closing line that ends on the result rather than a moral.
• A lint. A script that fails on the catalogued tells of AI writing, drawn from Wikipedia's "Signs of AI writing" [2] and the vale-ai-tells style package [3]: words like "delve", "pivotal" and "showcase", "not just X, but Y" constructions, "serves as" in place of "is", staged openers and summarising closers.
• Substack's engagement findings. A scrape of 94,391 posts [4] sets the targets the lint warns on: titles of 9 to 17 words (13 to 17 word titles averaged 26.8 reactions against 21.1 for 1 to 5 words), subtitles of 6 to 10 words, no question mark in the title, and an image (posts without a cover averaged 12.7 reactions, with one 22.9). Notes carry no links, since a link in a Note costs it reach [5].

Substack began scanning posts with Pangram on July 22, 2026. Any reader can score a post, reply or Note, and writers can test drafts, switch scanning off per post, or add a "How I make this" statement [6]. The launch drew complaints from writers worried about false accusations [7].

The ways people try to make model output read as human, strongest evidence first:

1. The human writes; the model edits. One widely shared rule: "You may not use a single word an LLM suggests to you", because "readers can detect LLM words in the parts per trillion" [8].
2. A voice profile with samples of the author. Epoch AI gave models five samples of a real author's writing and asked for new text in that voice. Pangram missed 10.10% of those passages, GPTZero 10.77% and Originality.ai 17.85%, against under 1% for plainly prompted text [9].
3. A humanizer pass, such as the blader/humanizer agent skill (about 52,000 GitHub stars), which rewrites a draft to remove 25 catalogued patterns [10].
4. Banned-word lists, the weakest: the vocabulary drifts with each model generation [2].

The skill took a middle path: the author supplies the substance, the model writes the prose in his voice. Option 1 had the best evidence; this was a bet that grounding and a voice profile would close most of the gap.

The first post covered two docs, Stylized shader looks and Atmosphere and dread. The angle: he is building a horror game and won't play a game with bad graphics; the research took him from no understanding to a working one; the surprise was how little a signature look needs, since most games commit to one or two effects.

The draft came out at about 300 words and passed the lint with no errors and no warnings. The author then edited it by hand in three places:

• He rewrote the title ("What makes a game look good, and what it means for my horror game").
• He cut a closing joke about his own game as "kind of forced".
• He swapped the closing question for one about horror.

Substack's check on the draft:

Fully AI-assisted text. AI 100%, AI-assisted 0%, Human 0%.

The author's own verdict on the joke he cut is worth keeping: he found the model's humour forced before any detector said anything.

What Pangram is

Pangram is not a list of tells. It is a neural classifier on top of an open-weight language model, with heads that label every token as human, AI-assisted or AI-generated; its report describes no hand-crafted linguistic features [11]. Its training data is built to defeat exactly the idea behind a voice profile:

• Synthetic mirrors. For each human document, a model is asked for its topic and then told "Write an article about X". The pair teaches the classifier "How was this text written?" rather than "What is this text about?" [11]. The original version of the method added hard negative mining: scanning tens of millions of human documents for the ones the classifier got wrong, and training on mirrors of those [12].
• AI-assisted examples come from human documents that a model has then edited [11].

Its reported robustness, on its authors' own tests [11]:

• Commercial humanizers: labelled AI 97.67% of the time, and AI or Mixed 98.83%.
• The blader/humanizer skill, applied by GPT-5.5, Opus 4.8 and Sonnet 4.6: 44 misses in 10,223 documents, a 0.430% false negative rate.
• Epoch's style-imitation set: 2.86% missed across its plain and style-imitating AI passages combined, with no human passage flagged. Epoch had measured 10.10% on style imitation alone for the Pangram version it tested [9].
• False positives: 0.0041%, about 1 in 24,000 human documents. An independent evaluation also found Pangram's false positives "essentially 0" across most thresholds [13].

Labels are assigned clause by clause. A document reads as Human when at least 90% of it is human, as AI when at least 80% is AI (AI-assisted clauses count at half weight), and as Mixed otherwise [11].

What the draft gave away

Pangram doesn't say which signals it learned, so this part is a reading of the draft against the author's real writing, not a description of the model.

• Predictable wording. Every sentence took the likely next word. "The answer was a lot smaller than I expected." The author's 2019 post has small oddities a model doesn't produce: "redirected-to HTTPS… so I got a 15/100", "Easy-enough." [1].
• Even rhythm. Every paragraph had the same shape: a setup, two or three supporting sentences, a payoff. The profile asked for short sentences, and the draft delivered them uniformly. Human writing is uneven: a long sentence, a three-word one, a tangent.
• Textbook moves. Three escalating fragments for effect ("Return of the Obra Dinn goes even further. Two colors and a pattern of dots. That's the whole look."), a glossing parenthesis ("(a small program that decides the color of each pixel)"), a quote with "as Thomas Grip put it", a closing question. Each is a move the model has seen work many times, and a classifier trained on its output has seen it many times too.
• No slack. Every sentence advanced the argument. Nothing repeated, nothing wandered. The author's own writing has slack: one paragraph of his 2019 post turns on "However, it tends to be technical in nature." and the next on "However, this was my project." [1]. A model would have varied the second.
• Rules applied as rules. The profile's habits ("Problem was:", the colon reveal) appeared exactly where the instructions put them. A person uses a habit when it occurs to them; a model following a profile uses it on schedule, which is itself a pattern.

The lint did its job: no AI vocabulary, no "not just X, but Y", no summarising closer, no em-dashes. It checks words and sentence shapes. Pangram judges the distribution the words came from, which no word list reaches. Worse, every rule added to a style guide or a lint gives the model one more thing to do consistently, and consistency is part of what a classifier learns.

The skill no longer writes the post. It follows the one approach with consistent evidence behind it [8]:

1. The skill outlines. It gives the story arc and, for each paragraph, its job, the facts and links from the docs, and a question for the author to answer. No finished sentences.
2. The author writes every sentence, in his own words.
3. The skill reviews only grammar, spelling and coherence, plus the facts against the docs and the site rules, and points out each issue rather than rewriting the sentence.
4. The skill suggests places for humour that fit the voice profile, and the kind of joke that would work there. The author writes the line, or skips it.

Pangram's clause-level labels are why the review stays light. A corrected typo leaves a clause his; a clause the model rewrote becomes AI-assisted, and enough of those move a post from Human to Mixed [11].

• One post, one detector. The result is a single data point against Pangram 4, which itself changes with each release.
• Section 4's second half is interpretation. Pangram's learned signals aren't published, so the reading of the draft is informed guesswork, checked only against the author's own writing.
• This doc was drafted by Claude Code from the working session and checked against the sources below. It is research for this site, not a Substack post, which is the kind of writing the new workflow keeps human.

1. [1]
   C. Sarkosh, "My experience getting an A+ from Mozilla’s Observatory tool on AWS," Medium, Sep. 15, 2019. Accessed: Sep. 24, 2026. [Online]. Available: https://medium.com/@csarkosh/my-experience-getting-an-a-from-mozillas-observatory-tool-on-aws-f0abf12811a1
2. [2]
   Wikipedia contributors, "Wikipedia:Signs of AI writing," Wikipedia. Accessed: Sep. 24, 2026. [Online]. Available: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
3. [3]
   tbhb, "vale-ai-tells," GitHub repository. Accessed: Sep. 24, 2026. [Online]. Available: https://github.com/tbhb/vale-ai-tells
4. [4]
   S. Günel and Orel, "How to Grow on Substack: 7 Data-Backed Tips From 94,000 Posts," Write Build Scale, Jun. 3, 2026. Accessed: Sep. 24, 2026. [Online]. Available: https://writebuildscale.substack.com/p/i-analyzed-94391-substack-posts-heres
5. [5]
   Orel, "How To Make URLs in Your Notes Convert Readers into Subscribers," The Writing Edge, Aug. 1, 2026. Accessed: Sep. 24, 2026. [Online]. Available: https://thewritingedge.substack.com/p/how-to-make-urls-in-your-notes-convert
6. [6]
   Substack Team, "How writers are reacting to Substack’s AI transparency tools," On Substack, Jul. 24, 2026. Accessed: Sep. 24, 2026. [Online]. Available: https://on.substack.com/p/how-writers-are-reacting-to-substacks
7. [7]
   J. C. Ofonagoro, "Substack AI Detector Sparks Writer Backlash," eWeek, Jul. 31, 2026. Accessed: Sep. 24, 2026. [Online]. Available: https://www.eweek.com/news/substack-ai-detector-writer-backlash/
8. [8]
   "How To Write With An LLM," A Final Ward (sockpuppet.org), Sep. 17, 2026. Accessed: Sep. 24, 2026. [Online]. Available: https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/
9. [9]
   J. Lee, "AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors," Epoch AI, Jul. 15, 2026. Accessed: Sep. 24, 2026. [Online]. Available: https://epoch.ai/data-insights/ai-detectors-false-negatives
10. [10]
    blader, "humanizer," GitHub repository, v3.0.0. Accessed: Sep. 24, 2026. [Online]. Available: https://github.com/blader/humanizer
11. [11]
    B. Glickenhaus et al., "Pangram 4 Technical Report," arXiv:2607.27183, Jul. 2026. Accessed: Sep. 24, 2026. [Online]. Available: https://arxiv.org/abs/2607.27183
12. [12]
    B. Emi and M. Spero, "Technical Report on the Pangram AI-Generated Text Classifier," arXiv:2402.14873, Feb. 2024. Accessed: Sep. 24, 2026. [Online]. Available: https://arxiv.org/abs/2402.14873
13. [13]
    M. Robinson, "Do AI Detectors Work Well Enough to Trust?," Chicago Booth Review, Dec. 2, 2025. Accessed: Sep. 24, 2026. [Online]. Available: https://www.chicagobooth.edu/review/do-ai-detectors-work-well-enough-trust