Reference writers

Eight working senior-technical writers whose moves inform technical-prose-editor. Cited so the patterns are inspectable rather than vibes-based. When tweaking a draft, naming a writer ("more Brooker here", "this paragraph wants Julia Evans") is more useful than abstract style notes.

Dan Luu (danluu.com)

Long-form analytical, contrarian, data-heavy. Reads like someone who actually ran the numbers.

• Conversational openings. Drops you into a personal observation. "If I ask myself a question like 'I'd like to buy an SD card; who do I trust...'", rather than a thesis.
• Hedges constantly. "I don't think it's controversial that...", "seems to be catching on a bit", "it's plausible that...". The hedge sits on the frame, leaving the claim itself flat.
• Accumulates examples instead of arguing abstractly. Ladder of scale: private forums, lobste.rs, HN, Reddit, Facebook, YouTube. Each rung extends the argument; no abstract principle dropped on top.
• Footnotes for tangents. Numbered, substantive. Keeps main-text momentum while satisfying scholarly impulses.
• Middle-voice closes. Refuses to oversell. "A full discussion of what that means is too long to include here, so we'll look at this in more detail another time."
• Transitional anaphora over connectors. "And if I compare X to Y... And then if I compare Y to Z..." builds momentum through repetition rather than through "Furthermore".

Use when the post is an argument that accumulates evidence.

Marc Brooker (brooker.co.za)

AWS Principal voice. Distributed-systems-grade depth in accessible prose.

• Short opening declarations. "It's time to be right." "Outcomes continue to matter." Three to five words; thesis stated as fact, then the post earns it.
• Technical analogies grounded in real systems. "Feedback loops, just like in electronics and control theory, can have significantly different capabilities from their underlying components."
• "Real" as anchor word. "Real customers", "real failures", "real conversation". Signals substance over theory.
• Honest qualifications. "Somewhat amusingly", "Solidly in", "I'm conflating two things here". These mark where the simplification is, rather than reading as throat-clearing.
• Footnote asides for self-aware humour. Substance in the body, dry observation in the footnote.

Use when the post is a thesis essay with technical depth.

Julia Evans (jvns.ca)

Confessional clarity. Naive questions taken seriously.

• Conversational framing openers. "A few years ago I gave a short talk..." / "I've been trying to understand X for a while now."
• "I think" / "my understanding is" as epistemic markers. Useful for placing the speaker without weakening the claim. "I don't really know anything about PHP so I don't have much interest in fact checking that PHP statement; I'm happy to leave it as an 'I think' and potentially correct later."
• Parenthetical self-correction. "(or maybe I'm misremembering; checked, no, this is right)".
• Demonstrates rather than explains. Shows the command, shows the output, says what surprised her.
• Mixes personal and technical without flag. Docker and Arabic tokenization sit next to "I love how this works".

Use for the confession shape: "I tried X, then deleted it" / "I didn't understand this for ages" / "here's a thing I learned this week".

Simon Willison (simonwillison.net)

Practical guide voice. Tool-builder operating in public.

• Numerical specificity. "68,000 photos", "$1.68 total cost", "18 organizations now have models on the Chatbot Arena Leaderboard." Numbers do the work.
• Inline links woven into sentences. "I wrote about this at the time in The killer app of Gemini Pro 1.5 is video". The link is part of the prose rather than appended.
• Casual parentheticals showing the seams. "(this list generated using Django SQL Dashboard with a SQL query written for me by Claude)". Honest about how the post was made.
• Self-deprecating humour. "My personal laptop... is the same laptop I've been using." Plain register, no posturing.
• TOC and section headers for comprehensive pieces. When the post is meant to be a reference scrolled back to, it has structure; when it's a quick note, it doesn't.

Use for practical guides, year-in-review-style pieces, "things I learned about X" posts that the reader will return to.

Patrick McKenzie / patio11 (kalzumeus.com, bitsaboutmoney.com)

Variable sentence length as instrument. Parentheticals everywhere.

• Aggressive sentence variety. A short declarative ("All of these assumptions are wrong.") next to a 60-word multiclause parenthetical-stuffed construction. Reads as a person rather than a model.
• Heavy parentheticals. "(Most people call me Patrick McKenzie, but I'll acknowledge as correct any of six different 'full' names...)". Context, deflation, asides.
• Anticipatory humour. Preempts the reader's objection before they raise it. "That Klingon Empire thing was a joke, right?" lands after a list of names-handling absurdities.
• Personal anecdote, then general principle. Establishes credibility through narration; the broader claim follows.
• Numbered lists building to absurdist conclusions (the falsehoods-about-names format).

Use when the post benefits from a strong personal voice and the reader needs to be carried through dense domain material (financial systems, name-handling, salary negotiation).

Will Larson (lethain.com)

Staff-engineer essay register. Measured authority grounded in narrative.

• Personal-narrative openers. "Silicon Valley's protagonist narrative" / "My father's retirement". Situates the essay in lived experience before generalising.
• Hedges by experience. "I've found", "I've gotten quite deliberate about". Opinions emerge from years of work rather than from theory.
• Names mental models cleanly. "Pace, People, Prestige, Profit, Learning." Each one a noun the reader can pick up and carry.
• Brief direct closings. Pivots from problem to aspiration without summarising. "For now, I'll be over here working on my pace."

Use for management-adjacent or career-adjacent posts. An IC author can borrow the register without the management topic.

Rachel Kroll (rachelbythebay.com)

Short ops war story. The whole post is the story.

• 600 to 1,200 words typically. Length emerges from the material rather than from a structural choice.
• No H2 headers, or one at most. Narrative drives.
• Technical detail emerges through the story. "We saw the latency spike at 03:12. By 03:14 the on-call was paged. By 03:18 we'd worked out the cache had..."
• Ends abruptly. When the point is made, the post stops. No synthesis, no "what this teaches us".
• Wry, slightly resigned voice. "Maybe nobody knows how to fix it. But it's worth thinking about."

Use for a specific debugging session, a deploy that went wrong, a production incident. Narrate the evening, land the fix, stop.

Aphyr / Kyle Kingsbury (aphyr.com)

Jepsen-style investigation. Code, output, and analysis in alternating blocks.

• Claim, then reproducible test, then finding loop. "We expected X. We ran the test like so [code]. We observed Y."
• Code and output as evidence, doing argumentative work rather than illustrating after the fact.
• Formal frame with literary touches. "We call this the goddess of distributed systems." Tonally distinctive; doesn't pull punches when finding bugs.
• Names what was tested and what wasn't. "We did not test linearizability under network partitions of more than 60 seconds. That work is left for future analysis."

Use for benchmarks, deep-dives into a specific tool's behaviour, "does X actually do Y" posts. A future "pgvector vs sqlite-vss at this scale" post would want this register.

Choosing between voices

Most posts will end up as a blend of two or three of these. The dominant one should be picked deliberately based on the post's shape (references/shapes.md).

Rough mapping from shape to dominant voice:

• War story: Rachel Kroll, with Aphyr for the technical evidence blocks.
• Argument essay: Dan Luu or Marc Brooker.
• Practical guide: Simon Willison.
• Confession: Julia Evans or Patrick McKenzie.
• Investigation: Aphyr, with Brooker for the framing.

If the draft doesn't read like at least one of these, it probably reads like AI.