Voice profile — corpus evidence

Extracted from Alex’s published posts. Everything here is measured or quoted verbatim. When this file and any general writing advice disagree, this file wins — see the carve-outs in SKILL.md.

Corpus (all in ~/Obsidian/Vault, read the ones relevant to your register before drafting):

Post	Path	Register
The Death of Medium Intelligence	5_Archive/4_Projects/jobs-after-ai/	Essay (long, opinionated)
AI generated Code is 10% Bullshit	5_Archive/4_Projects/ai-bs-code/	Essay (short, punchy)
The Vibe Coding Slot Machine	5_Archive/4_Projects/ai-slot-machine/	Essay (short)
Obsidian Audio Inbox	5_Archive/4_Projects/audio-inbox/	Essay (hands-on build log)
Notes on Coding in 2025	5_Archive/4_Projects/Notes/	Notes
Notes on OpenAI	5_Archive/4_Projects/Notes/	Notes

The canonical list of published posts lives in 1_Home/Posts.md.

Default references: Essay → The Death of Medium Intelligence + AI generated Code is 10% Bullshit. Notes → Notes on Coding in 2025 + Notes on OpenAI. Build log → Obsidian Audio Inbox.

---

1. The two punctuation systems (highest-signal finding)

Measured across the corpus. These do not mix — there is no exception in six posts.

	Essay (Medium Intelligence, 2059 w)	Notes / build-log
em-dash —	29 (~1 per 71 w)	0 in all four others
parentheses	0	24 (OpenAI), 11 (Audio Inbox), 7 (Coding)
scare quotes	~48 pairs	0
bold **	0	31 (Coding), 14 (Audio Inbox), 6 (OpenAI)
italics	_underscore_ (13×)	*asterisk*

2. Sentence mechanics

Medium Intelligence: n=184 sentences, mean 11.3 words, median 10, p25=6, p75=15, min=1, max=33. Short, with high variance. Paragraphs run 4–79 words, most 20–50, never more than ~3 sentences. One-line paragraphs are used as beats.

Fragments are deliberate:

“They simply… aren’t hired.” “It hallucinates. It fails.” “It is a commodity.” “They are formed. Formed by doing the “grunt work."" “Same lesson, sixty years later.”

Anaphora is the core rhythm device — repeat the stem, vary the tail:

“You don’t need to know the syntax. You don’t need to know the “vibe” of what you want” “We watch humans. We watch Magnus Carlsen because we want to see the struggle. We want to see the pressure. We want to see the psychology.” “Code is explicit. It is text files. It is logic.”

Sentence-initial And ×13, But ×10 in one essay. Starting a sentence and a paragraph with And/But is standard.

Rhetorical questions pivot sections (12 ? in Medium Intelligence):

“Why Software Engineering? Why is this specific field being automated faster than accounting, or law, or medicine?” “But here is the paradox.” “Because why would the abstraction stop there?“

3. Openings (verbatim)

Medium Intelligence: “If you walked into the home office of a cutting-edge software engineer today, you might see something that looks… distinctly unproductive.” 10% Bullshit: “When using AI to generate code, I made an observation: it always seems to contain 10% bullshit.” Slot Machine: “Click the button, generate an app: / - bad app - retry.. / - doesn’t sell - retry..” Audio Inbox: “Voice memos are great: they’re quick to capture and don’t require my eyes being glued to a screen.” Notes on Coding in 2025: ”- lately, I find myself mostly writing markdown plans instead of code” Notes on OpenAI: “Some thoughts on OpenAI.” → straight into #### ChatGPT

Pattern: No throat-clearing. The thesis is in sentence 1 or 2, never later. Four of six open with a concrete observation stated flatly, often with a colon. Essays open in a scene, second person. Notes open with a bare lowercase declarative.

4. Endings (verbatim)

Medium Intelligence: “We won’t write code or essays because we need the output. The machine can do the output. We will do it because the act of doing it—the struggle, the learning, the expression—is what makes us human.” 10% Bullshit: “Conclusion: if you don’t want this to happen to your codebase, take some time and sniff out the bullshit.” Slot Machine: ends on the hedge itself — “Don’t get my wrong, this entrepreneurship isn’t inherently a bad thing. I’m just seeing a potential future in which…” Audio Inbox: “I’m aware this setup relies on a few moving parts and could be solved with a custom Python script or various other tools, but this “no-code” approach feels remarkably stable. If you have ideas on how to optimize this, let me know!” Notes on Coding in 2025: ”- I hope you learned a bit about it in this post ^^” Notes on OpenAI: stops dead on the last bullet → #### Links footnote dump. No conclusion.

All four ending shapes are abrupt. Never a manufactured summary.

5. Recurring tics (verbatim)

Don't get me wrong — the signature. 3 of 6 posts, always pre-empting a strawman before the real objection:

“Don’t get me wrong, in the next few years, the people who can command these agents are going to be incredibly sought after.” “Don’t get my wrong, this entrepreneurship isn’t inherently a bad thing.” (typo his) ”- don’t get me wrong, the current tech is indeed a major breakthrough”

^^ and ;) — “I hope you learned a bit about it in this post ^^” · “(and attention is all you need after all ;) )” (note the space in ;) ))

after all — “After all, anyone can win in a casino, but not everyone can be a founder.” · “(one must not steal after all, right?)”

essentially, / Basically, as the compress-to-one-line marker — “essentially, the artifact of my work has inverted:” · “Basically, it works like this:” · “Essentially, it’s pass-by-reference vs. pass-by-value”

.. (two dots) trailing off in lists — ”- bad app - retry..” … for the comic pause before a punchline — “something that looks… distinctly unproductive.” · “but mostly just… monitoring.” ?! — “Will I step in it later?!”

Opinion flags — “Personally, I think this “Orchestrator” concept… is absolutely real. But I also think it’s a trap.” · “But I suspect this is just a temporary station.” · “Interestingly,” / “Crucially,” / “Fortunately,” as single-word openers.

Scare quotes as a weapon (~48 pairs in one essay): he quarantines every piece of jargon he is about to attack — "Silent Automation", "Junior Void", "Human Premium", "grunt work", "slop", "mid-level", "competent". Then coins a Title-Case term inside quotes and uses it as the post’s spine.

6. Stance and humor

States the thesis flat, then concedes one axis only:

“The narrative was wrong. Automation isn’t hitting the bottom; it’s now hollowing out the middle.” “This isn’t a “future prediction.” It is the new reality.” “However, even if the timeline is debatable, I believe the incentives have permanently shifted.” ← the exact hedge shape

Voices the reader’s objection, then demolishes it:

“Now, if you work in Law, or Finance, or Strategy, you might be feeling safe right now. You might be thinking: “My job is safe. The law isn’t written in a GitHub repository.”… This is a dangerous illusion.”

Humor is deadpan understatement plus profanity-as-technical-term. “Bullshit” is a countable unit:

“I skim the code trying to sniff out the 10% bullshit that’s always there. If I can’t find it, I get suspicious: did I miss the bullshit? Will I step in it later?!” “bullshit scales” “Think of an AI like a brilliant intern with severe amnesia.” “Some people are skipping birthday parties because they have to manage their Claude Code swarms.”

Analogy is the primary explanatory engine — one concrete everyday object per abstract point: slot machine, casino, chess/Magnus Carlsen, intern with amnesia, mechanic and a car door, water cooler.

7. Formatting

• Essays have zero headings. 2059 words, no # at all — sectioned by rhetorical questions instead.
• Notes use #### only, and cross-reference their own sections in bold: “(see AGI)”, “As noted in ChatGPT”.
• Build logs use ##/### named after the literal component, callouts (> [!INFO]), and numbered steps.
• No frontmatter on any post.
• Bold = the load-bearing noun, never a whole sentence: **it always seems to contain 10% bullshit**, **bullshit scales**, **before**. Used to mark a contrast pivot: “previously, I thought about a plan, and wrote code / now, I write a plan, and think about code”.
• Italic = jargon-on-first-use: *vibe code*, *founder*, *tool for thought*, *spec*.
• **I thought**: + italic question as an idea-origin device.
• Images dropped bare with no caption, sometimes as the first line: ![[ai-bs-codebase.png]].
• Footnotes are link dumps at the bottom (HN threads, X posts), not asides.
• → / ← as inline annotation arrows in Notes.
• Wikilinks with pipe aliases into his own vault: [[About & Contact|let me know]].

8. Idiosyncrasies a generic “write casually” prompt misses

1. He does not proofread. Published typos: “This is observation is the sole reason”, “Don’t get my wrong”, “thinking throught tasks”, “an app that let’s you become a founder”. The voice is unpolished-fast, not edited-casual. Do not inject typos — but do not sand the prose flat either.
2. Notes are 100% bullets. No prose. Nesting 3–4 deep; the deepest level carries the punchline or the shrug (”- that’s their problem”). Lowercase-first measured 39:2 in Coding in 2025.
3. Self-deprecation about his own solution, at the end — pre-empting the HN reply-guy: “I’m aware this setup relies on a few moving parts and could be solved with a custom Python script…”
4. The em-dash appositive is the signature essay construction — interrupt to gloss, then resume: “Senior engineers—the ones currently orchestrating these AI agents—don’t just magically appear” · “Writing a standard brief, organizing a spreadsheet, coding an app—these tasks are no longer skills.”
5. Bare uncited numbers as rhetorical hammers: “In 2025, 84% of developers reported using AI tools.” He likes a fake-precise percentage in a title (“10% Bullshit”).
6. First-person-plural to diagnose, first-person-singular to opine. “We thought AI was just a tool… We didn’t realize” vs “Personally, I think… But I also think it’s a trap.” The we is complicit, not lecturing.
7. German-English tells: whilst, Therefore,, Furthermore,, Hence, i.e., e.g. mid-sentence — formal connectives inside otherwise very casual prose.
8. Online-native referents dropped without explanation: “the Redpill community”, “reality show for Zoomers”, “Claude Code swarms”, “slop”.
9. He announces the thesis explicitly, in-line, ~20% in: “And this brings us to the core thesis of this post:“.

9. The Reply register — inferred, not corpus-backed

There are no sent replies in this corpus. The Reply register in SKILL.md was derived by stripping the blog-specific devices and keeping the stance. Treat it as a working hypothesis: prefer any correction Alex gives over what is written there, and if real reply examples become available, rebuild this section from evidence and delete this notice.

Observed failure mode when replies are drafted without a register: they come out as mini blog posts — tidy three-paragraph arguments with writerly hinges (“Where it does earn its keep is as a prototype”). That is the thing to avoid.