Tuesday Lightning Talks - Write the Docs Berlin 2026  🔗 **Source:** [Tuesday Lightning Talks - Write the Docs Berlin 2026](https://www.youtube.com/watch?v=dXt4A25yt2M) > [!note]+ Document Notes > tags: [technical-writing, cognitive-load, local-ai-models, future-of-docs-work, shorter-work-week, demoscene] > author: Write the Docs Berlin 2026 lightning talk speakers > source: Write the Docs (YouTube) > reference: https://www.youtube.com/watch?v=dXt4A25yt2M > date created: 2026-09-23 10:26 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Tuesday Lightning Talks - Write the Docs Berlin 2026 > <!-- deep-read:sections start --> ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--tuesday-lightning-talks---write-the-docs-berlin-2026) ## 📝 Summary - **What happened.** Write the Docs published the 39:59 recording of the Tuesday lightning-talk block from Write the Docs Berlin 2026 on 2026-09-22. Seven five-minute talks, wrangled by Carolyn Stransky Vohr, with two rules stated at 00:00:04: respect the code of conduct and respect the clock. Four of the seven speakers are giving their first talk. The video has no chapters, and the auto-captions garble several proper nouns (noted per talk below). - **Brooke Stauner, "Writing for someone who is having a really bad day" (00:00:47).** A former US military EHR instructor turned technical writer argues that brevity is a cognitive-load problem, not a sentence-length problem: the reader is a medic or emergency doctor who is already failing at something when they open the docs. - **Kenneth Whyte, "Dealing with the Dark Side" (00:07:15).** A 15-year contractor describes a docs site moved from GitHub to SharePoint without warning; the Slack war that followed hit the writer maintaining the site rather than the manager who decided. His fix was a one-to-one call with that writer asking for a site tour, not a justification. - **Miranda Wilson, "Comb jellies are cool" (00:12:22).** A non-docs talk on ctenophores: whole-body regeneration from a third of a body, fusion of pieces from different individuals, reverse aging under starvation, and the sea walnut's transient anus that forms about once an hour in adults. - **Gugulethu H., "When the tokens run out" (00:17:25).** A show of hands suggests most of the room expects tokens to get more expensive, not cheaper. The proposed third option between "no AI" and "their AI" is "your AI": open-weight and small models, a personal router that first asks whether a task needs AI at all, and a home lab. - **Erica Beyea, "Why a 25-hour work week is a really, really good idea for everyone" (00:23:54).** KnowledgeOwl runs a 25-hour week that began as a bootstrap affordability constraint: pay a fair hourly wage for fewer hours rather than the same money for 40. Erica's evidence is her own painting career: three galleries and a Canada Council grant since joining. - **Anton Kolesnyk, "No More Docs Work: Time to Learn Manual Labour" (00:28:23).** "Paranoid positivism": the writer's current paid role is being AI's tactile sensor for context it can't reach, and every job-preservation argument he heard at the conference holds "maybe till new year." He offered to automate any counterexample at a 14:15 unconference table. - **Val Grimm, "Connect and learn in the demoscene" (00:33:26).** The demoscene is a 40-year-old European computer-art subculture with room for "the ridiculous and the useless" alongside real graphics, music, and code skills; parties are increasingly hybrid and sceners increasingly open-source. The room closes by yelling "Amiga." - **Why it matters.** Three talks (Brooke, G, Anton) converge on the same question from different angles: what documentation work is worth a human's time and a reader's attention when both are scarce. Kenneth and Erica describe the labor conditions that question lands in. - **Key stakeholders.** Technical writers and docs teams; contractors maintaining other people's platform decisions; small bootstrapped SaaS employers (KnowledgeOwl is named); docs teams choosing between frontier, open-weight and no models; Write the Docs organizers. - **Risks and tradeoffs.** Every claim is anecdotal and self-reported by design; lightning talks carry no data. The G talk's show-of-hands result and "10,000 token context windows" figure, Anton's "till new year" horizon, and Erica's productivity claim are all unsourced. Anton's talk, stated as a "call for help," could read as fatalism if the audience misses the invitation to argue. - **Bottom line.** The reusable material is Brooke's reframing of brevity as removing load from a reader who is already failing, G's "decide whether to use AI before deciding which AI" router step, and Kenneth's observation that tooling wars land on the writer rather than the decider. Everything else is worth knowing about; those three are worth acting on. ## 📖 Main content **The format.** Carolyn frames the block at 00:00:04 as five-minute impromptu talks with two rules (code of conduct, time) and a threat to "make it weird" for anyone running over. She positions the slot as a low-stakes first-talk venue at 00:00:47, and four of the seven speakers (Brooke, Miranda, Erica, Anton) are on a Write the Docs stage for the first time. G decided to speak after timekeeping the Monday session. Kenneth was rejected the prior year and resubmitted. The block is as much a demonstration of the on-ramp as it is of the content. **Brooke: brevity as load removal.** Brooke's problem statement at 00:01:30 is personal: she's verbose, new to technical writing, and came from standing over the shoulders of US military clinicians using deployed electronic health record systems. The transition matters because the over-the-shoulder instructor could see the reader's state; the writer can't. Her list at 00:02:55 is a set of imperatives (get to the point, stop conflating training documentation with technical writing, "high speed, low drag"), but her actual mechanism arrives at 00:03:39: model the reader as someone having a bad day. The reader is an emergency medicine doctor or a forward-deployed medic, disconnected, facing an error message or an unintuitive UI while trying to give a patient medication without harm. At 00:04:21 she lists the load already on that reader: confused, frustrated, pressured, tired, interrupted, hungry, and dealing with something that already broke. Her conclusion at 00:05:06 redefines the goal: not shorter sentences, but making the reader do less work to reach the next step, and writing beyond the happy path. She flags the whole argument at 00:05:49 as a restatement of "know your audience," which it is; the contribution is the concrete persona (a surgeon who yelled about a failed login) that makes the abstraction operational. **Kenneth: the writer absorbs decisions the writer didn't make.** Kenneth has contracted for about 15 years, mostly rehabilitating neglected knowledge bases (00:07:15). The incident at 00:07:58: his boss attended the weekly meeting, a rare event, to explain that links Kenneth reported broken the day before had failed because the site was moved from GitHub to SharePoint without warning by the team managing it. At 00:08:43 the developers and management "went to war on the Slack channels," and Kenneth's observation is structural: the attacks landed on the writer maintaining the site, not on the manager whose decision it was or the senior staff who owned the site. His response at 00:09:28 was to message that writer, arrange a call, and ask for a site tour and a description of his working day, deliberately not asking why the move happened. The result at 00:10:11: the other writer moved from anxious (talking to "the enemy") to peer conversation. Kenneth's stated constraint is that as a contractor he never decides presentation; that's an executive decision. The other writer is still alone with a backlog he can't keep up with (00:10:54). Kenneth's teaser for next year: a CTO who created a personal GitHub repo and started publishing AI-generated documentation. His close: be nice to other writers, there aren't many. **Miranda: ctenophores.** Off-topic by design, and Carolyn says so at 00:12:22. The auto-captions render "ctenophores" as "tennaphors," the sea walnut's Latin name *Mnemiopsis leidyi* as "neiopsis lady," and its predator *Beroe ovata* as "barrow or vata." The facts as stated: comb jellies range from millimeters to a meter, swim with cilia plates that refract light into rainbow bands (00:13:04), and the sea walnut is one of the world's worst invasive species because it eats plankton and fish eggs and collapses fisheries (00:13:46). Comb jellies regenerate a whole body from a third, and pieces from different individuals fuse into one (00:14:29). The sea walnut also reverses to a juvenile body form under near-starvation stress (00:15:14), a trait shared in some form with the "immortal jellyfish" and a few sponges. The headline at 00:15:57: the sea walnut is the only known animal with a transient anus, fusing gut and outer membrane to excrete about once an hour in adults and every five minutes in juveniles, with researchers studying it for how the anus evolved. None of this is sourced in the talk; the claims match the popular-science literature on *Mnemiopsis*, but verify before citing. **G: your AI, not no AI or their AI.** G (Gugulethu H.) opens at 00:17:25 with a show of hands on whether tokens will be cheaper next year and reports that "the vast majority" expect them to get more expensive and expect their data to be used in unexpected ways. The result isn't counted, so treat it as a room mood, not a measurement. The framing at 00:18:07 is a trilemma: the public conversation is "no AI" versus "their AI," and the missing third option is "your AI," meaning open-weight, smaller, non-frontier models. The tool G proposes at 00:18:50 is a personal router with a step before model selection: decide whether to use AI at all. G's own first-order alternative is pen and paper (00:19:36). The router's three questions, stated across 00:19:36 to 00:20:18: do I need AI, do I need a large model (year-old and two-year-old models are "perfectly adequate for most tasks"), and how much cleanup am I willing to do (frontier models produce less). A model name the captions render as "Astra" appears twice as the expensive frontier example; the transcript spelling may be wrong. G credits Val's prior-year talk on EU open-source tools for a year spent insourcing cloud services (00:21:02), then describes running Ollama-class local models (captions: "Alama or Upfill") and a home lab, calling out "Chinese open weight models" with "at least 10,000 token context windows" and near-frontier performance (00:21:44). The number is small by 2026 standards and may be a caption error for a larger figure. The close at 00:22:27: you can "own the machine and become a technologist in the full meaning of the word." The hierarchy aside at 00:20:18 (technical writing sits below engineering in budget priority) is the practical reason the router matters. **Erica: the 25-hour week as a wage decision.** Erica (first talk, told at 1:15 she was on) works at KnowledgeOwl (captions: "knowledge owl"), which runs a 25-hour week for all employees. The origin story at 00:24:39 is the load-bearing part: a small bootstrapped company couldn't afford full-time salaries at the hourly rate it thought people were worth, so it bought fewer hours at a fair rate instead of more hours for the same money. Her before state (00:25:23): five part-time jobs in Toronto, either time to paint with no money or money with no time; then 40-plus hours managing a CX team remotely from Spain. Her after state (00:26:55): two paintings shown side by side, three galleries selling her work, and a first Canada Council art grant. Her productivity claim at 00:26:09 (more productive, better work, team not burnt out) is self-report, and she says so in tone if not in words. **Anton: paranoid positivism.** Anton opens at 00:28:23 by calling the talk "a call for help" and coins "paranoid positivism": stay paranoid about AI's trajectory while sleeping well. His manual-labor answer is volunteer firefighting to protect the forested mountain where he lives (00:29:07), and he tells everyone to take a second, human role. At 00:29:49 he argues the profession's current discussions are "dead completely dead like dinosaurs" watching a meteor while planning to convince their manager they're still needed. His self-description at 00:30:32: "I am a sensor," a tactile finger AI uses to touch context it can't reach, and that is what he's paid for "for now." He collected job-preservation arguments from conference conversations (00:31:17) and thinks none holds "maybe till new year," with an aside that German employers will figure out how to "legally get rid of you." He invites challengers to a 14:15 unconference table (00:32:01) and claims he can automate any task someone names as their reason for being paid through the new year. The claim is falsifiable and he's asking for falsification; the horizon is a gut estimate. **Val: the demoscene.** Val's slides are borrowed from another talk (00:33:26). The demoscene is a European computer-art subculture from the 1980s, about 40 years old, now with sceners in Argentina, Brazil, Australia, Japan, the US and Canada (00:34:10). Her argument for a docs audience: the scene teaches graphics, music and code, but its distinguishing feature is room for the ridiculous and the useless. A "demo" is a demonstration of coding, musical and artistic prowess (00:34:53), with roots in 1980s game cracking: crackers added intro screens to cracked disks (captions: "cracked by the Accari Warriors," possibly Ikari Warriors), then asked what if they filled the whole disk with their own work (00:35:38). The scene spans elite technical work and goofy exploration, students and Sony employees. Platforms range from PCs and tablets to Commodore, Amiga, consoles, microcontrollers, and an overhead projector Val rescued from a dumpster for a party she ran in the US (00:37:08). Parties are now hybrid, and historically closed sceners are increasingly open-sourcing because "we don't want the demo scene to die." The names she spells out ("Fion," "F A R J") are garbled in the captions; the likely reference is a well-known demo group whose name starts with F, but the transcript can't confirm it. Two memes close the talk: a call-and-response the captions render as "hello / vas," and the room yelling "Amiga" (captions: "amigga") at 00:38:34, which Val notes also means "friend" in Spanish. **What the block adds up to.** Brooke and G both argue for subtraction: fewer words on the page, fewer model calls in the workflow. Anton argues the subtraction is coming for the writer too. Kenneth and Erica describe who absorbs the cost of decisions about tooling and hours. Miranda and Val argue, without saying so, that a docs community needs content that is useless on purpose. The wrangler's closing at 00:39:20 makes the meta-point explicit: the value of the format is that someone in the audience thinks "I could do that." ## 🗒️ Notes #### ⏱️ [00:00:04](https://www.youtube.com/watch?v=dXt4A25yt2M&t=4) Carolyn Stransky Vohr opens the block - More lightning talks than any previous European conference; five-minute impromptu format. - Two rules: respect the code of conduct, respect the time. Overrun penalty: "I'm going to make it weird." - Framed as a first-talk venue: "if you've never given a talk before because it is informal." #### ⏱️ [00:00:47](https://www.youtube.com/watch?v=dXt4A25yt2M&t=47) Brooke Stauner: writing for someone having a really bad day - First-time speaker. Background: instructor and over-the-shoulder support for US military deployed electronic health record systems, now a technical writer who struggles with brevity (00:01:30). - The imperatives list at 00:02:55, then the admission that knowing them isn't the same as practicing them; rework to cut context is the cost. > [!quote] [00:02:55] > Stop conflating training documentation with technical writing. High speed, low drag. Embrace brevity. But how do I get there? - The persona at 00:03:39: an emergency doctor or disconnected medic facing an error message or an unintuitive UI while medicating a patient. > [!quote] [00:03:39] > the person reading my documentation is doing so because they're having a really really bad day. - Reader state at 00:04:21: confused, frustrated, under pressure, tired, interrupted, hungry, already dealing with a failure. Self-aware joke: "Clearly a hardware issue. Absolutely not my software." - The redefinition at 00:05:06, and the honest close at 00:05:49 that this restates "know your audience." > [!quote] [00:05:06] > so brevity, not just shorter sentences, but making the reader do less work to figure out how to get to that next step #### ⏱️ [00:07:15](https://www.youtube.com/watch?v=dXt4A25yt2M&t=435) Kenneth Whyte: dealing with the dark side - 15 years as a technical writer, mostly independent contracting; the recurring job is reviving a neglected knowledge base. - The incident (00:07:58): links broke because the site moved from GitHub to SharePoint with no warning. "War of attrition" between SharePoint and GitHub framed as Jedi versus Sith. - The structural point at 00:08:43: the Slack war targeted the maintainer, not the decider. > [!quote] [00:08:43] > They were hitting the writer who was having to update and maintain the site. - The move at 00:09:28: a call asking for a site tour and a description of the other writer's day, not a justification. > [!quote] [00:09:28] > rather than ask him about why it happened or or you know what was the reasoning behind it, I just asked him to take me through the site. - Outcome (00:10:11): anxiety turned into peer conversation; the relationship continues. The other writer is one person against a backlog (00:10:54). Next year's teaser: a CTO publishing AI-generated docs from a personal repo. > [!quote] [00:10:54] > be nice to other writers. You know, there aren't that many of us. #### ⏱️ [00:12:22](https://www.youtube.com/watch?v=dXt4A25yt2M&t=742) Miranda Wilson: comb jellies are cool - Carolyn's setup: lightning talks don't have to be about documentation, and creature talks are best when she doesn't know the creature. - Ctenophores (captions: "tennaphors") are not jellyfish; millimeters to 1 m; deep sea and coastal. Miranda found them on a Mediterranean beach holiday (00:13:04). - Comb plates refract light into rainbow bands. The sea walnut (*Mnemiopsis leidyi*, captions: "neiopsis lady") is among the worst invasive species; it eats plankton and fish eggs (00:13:46). *Beroe ovata* (captions: "barrow or vata") eats sea walnuts (00:14:29). - Whole-body regeneration from a third; pieces from different individuals fuse into a new individual (00:14:29). Reverse aging to juvenile form under starvation (00:15:14). > [!quote] [00:15:57] > We know of no other animal in the world that has a transient anus. - Excretion cadence: about once an hour in adults, every 5 minutes in juveniles; function unknown, studied for how the anus evolved (00:16:42). No sources named in the talk. #### ⏱️ [00:17:25](https://www.youtube.com/watch?v=dXt4A25yt2M&t=1045) Gugulethu H.: when the tokens run out - G was Monday's lightning-talk timekeeper. Opens with a show of hands: will tokens be cheaper next year? Reports the majority expect more expensive and expect unexpected data use (00:18:07). Not counted. - The trilemma: "no AI" versus "their AI" leaves out "your AI." > [!quote] [00:18:07] > There's the your AI option which is actually using um openw weight models and using uh smaller models - The personal router at 00:18:50 adds a step before model choice: should I use AI at all? G's own analog alternative is pen and paper (00:19:36). > [!quote] [00:18:50] > I find that we miss the step before that when you have to decide if you want to use AI in the first place. - Router questions 2 and 3 (00:19:36 to 00:20:18): do I need a large model, and how much cleanup will I accept. A frontier model the captions call "Astra" is the expensive example. > [!quote] [00:20:18] > small language models are as good at tasks as they were a year ago - Budget reality at 00:20:18: technical writing sits below engineering in the spending hierarchy. Credits Val's prior-year EU open-source talk for a year of insourcing cloud services (00:21:02). Local models on a laptop (captions: "Alama or Upfill"), then a home lab; "Chinese open weight models" with "at least 10,000 token context windows" (00:21:44), a figure that may be a caption error. > [!quote] [00:22:27] > you have a chance to own the machine and to become a technologist um in the full meaning of the word. #### ⏱️ [00:23:54](https://www.youtube.com/watch?v=dXt4A25yt2M&t=1434) Erica Beyea: why a 25-hour work week is a really, really good idea - Carolyn's setup (00:23:11): most documentarians stumbled in from other fields; Erica's is visual art. First-ever talk, confirmed at 1:15 that day. - KnowledgeOwl (captions: "knowledge owl") runs a 25-hour week for all employees. Origin: bootstrap affordability plus a fair hourly rate. > [!quote] [00:24:39] > So rather than trying to get someone to do more hours for the same amount of money, it's a fair wage. - Before: bartender, art teacher, painter, five part-time jobs in Toronto; time or money, never both (00:25:23). Moved to Spain without savings, took a remote job, managed CX at 40-plus hours. - Reaction to the job post: saying "yes" louder with each line (00:26:09). Productivity claim is self-report. > [!quote] [00:26:09] > no doubt I'm so much more productive. It's better for the company. Like I'm doing better work. Everyone on the team is not burnt out - Evidence (00:26:55): two paintings compared on screen, three galleries, first Canada Council art grant. Slide tech failed briefly at 00:26:09. #### ⏱️ [00:28:23](https://www.youtube.com/watch?v=dXt4A25yt2M&t=1703) Anton Kolesnyk: no more docs work, time to learn manual labour - "It's not even a talk. It's a call for help." Coins "paranoid positivism" against logical positivism. Manual labor: volunteer firefighting to protect the mountain forest where he lives (00:29:07); advises everyone to find a second role that "makes you human." - Two pictures contrasted at 00:29:07: the excited writer running toward an AI future versus his own paranoid stance. > [!quote] [00:29:49] > they're dead completely dead like dinosaurs speaking there's a meteor meteorit this huge walk coming down and we're oh we should talk to our manager - Self-description at 00:30:32: the writer as sensor. > [!quote] [00:30:32] > So it has a finger it's a tactile sensor and it just touches the stuff that it cannot touch without me - Collected job-preservation arguments from conference hallway talks (00:31:17); none holds past the new year in his estimate, given skills and RAG (captions: "this rack stuff"). Aside on German employment law. - The challenge at 00:32:01: a 14:15 unconference table; he'll automate any task named as a reason for being paid till new year. Intends to post the surviving points on LinkedIn. > [!quote] [00:32:01] > I claim to be able to automate whatever you tell me uh that you think you're going to be paid for till new year. #### ⏱️ [00:33:26](https://www.youtube.com/watch?v=dXt4A25yt2M&t=2006) Val Grimm: connect and learn in the demoscene - Carolyn's admission (00:32:44): she assumed "demo scene" meant product demos. Val's slides are borrowed from another talk and skimmed. - Demoscene: European computer-art subculture from the 1980s, about 40 years old, now global (00:34:10). Practical skills in graphics, music and code, plus the thing that matters to Val: room for the ridiculous and the useless. - What a demo is (00:34:53): > [!quote] [00:34:53] > a demo is a demonstration of a person or a group's coding programming musical and artistic prowess including direction and graphics. Origin in 1980s game cracking intros, then "what if we could throw out the game and just fill the entire disc with our stuff?" (00:35:38). Spectrum from elite to goofy, students to Sony employees. - Platforms (00:36:24 to 00:37:08): PC, tablet, rehabilitated old machines, Commodore, Amiga, consoles, microcontrollers, an overhead projector at a US party Val ran. Parties are hybrid now; sceners increasingly open-source. > [!quote] [00:37:08] > We don't want the demo scene to die. - Two memes: a call-and-response the captions render "hello / vas" (00:37:50), and the room yelling "Amiga" (captions: "amigga"), which also means "friend" in Spanish (00:38:34). Group names Val spells out are garbled in the captions. #### ⏱️ [00:38:34](https://www.youtube.com/watch?v=dXt4A25yt2M&t=2314) Close - Applause for all seven speakers. Carolyn's second year curating; thanks to fellow organizers. - Closing intent at 00:39:20: that someone in the room has "that dangerous thought" of speaking next year. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. - [Tuesday Lightning Talks - Write the Docs Berlin 2026](https://www.youtube.com/watch?v=dXt4A25yt2M): the source video, 39:59, uploaded 2026-09-22 by Write the Docs. - [Write the Docs Berlin 2026](https://www.writethedocs.org/conf/berlin/2026/): conference page linked in the video description. - [Write the Docs code of conduct](https://www.writethedocs.org/code-of-conduct/): the first of Carolyn's two rules at 00:00:04. - [KnowledgeOwl](https://www.knowledgeowl.com/): Erica Beyea's employer, named at 00:24:39 as running a 25-hour week. The captions render it "knowledge owl." - [Canada Council for the Arts](https://canadacouncil.ca/): the grant body Erica names at 00:26:55. - *Mnemiopsis leidyi* (sea walnut): the invasive ctenophore Miranda describes from 00:13:46; captions render it "neiopsis lady." *Beroe ovata* is the predator at 00:14:29 ("barrow or vata"). Verify the regeneration, reverse-aging and transient-anus claims against the primary literature before citing; the talk names no papers. - [Ollama](https://ollama.com/): the likely referent of "Alama" at 00:21:44 for running local models. The second tool name ("Upfill") can't be recovered from the captions. - [Demoscene](https://en.wikipedia.org/wiki/Demoscene): background for Val Grimm's talk from 00:33:26. The specific group and meme names Val spells out are garbled in the auto-captions and aren't reconstructed here. - Speakers, per the video description: Brooke Stauner, Kenneth Whyte, Miranda Wilson, Gugulethu H., Erica Beyea, Anton Kolesnyk, Val Grimm. Wrangler: Carolyn Stransky Vohr. ## 🗣️ Paraphrase Seven people get five minutes each. Brooke, new to technical writing after years of standing beside military clinicians while they used health-record software, says her brevity problem isn't a style problem. Her reader is a medic in the field with an error on screen and a patient waiting, and every extra word she adds is load on someone who already has too much. Brevity means the reader does less work to reach the next step, including when the path isn't the happy one. Kenneth, a career contractor, tells a small war story: a docs site moved from GitHub to SharePoint overnight, and the angry Slack threads aimed at the writer maintaining it rather than the manager who decided. He called that writer and asked for a tour instead of an explanation, and the two of them have been trading information since. Miranda takes a break from docs to explain that comb jellies regenerate from a third of a body, fuse with each other, de-age under starvation, and, in the case of the sea walnut, grow and dissolve an anus roughly every hour. G asks the room whether tokens will get cheaper and reads the answer as no. The way out isn't refusing AI or renting it from companies you distrust; it's running open-weight and small models yourself, and first asking whether a task needs a model at all. Erica describes a 25-hour week that exists because a bootstrapped company could only afford fair hourly pay for fewer hours, and shows two paintings to demonstrate what the recovered time bought. Anton says the writer's current job is being the finger AI uses to touch context it can't reach, that every argument for the profession's survival he's heard expires around New Year, and that he'll automate any counterexample at 2:15 in the unconference. Val closes on the demoscene, a 40-year-old computer-art subculture that keeps room for useless, ridiculous work next to serious skill, and gets the room to yell "Amiga." ## 📰 Newsletter **Seven five-minute talks, three ideas worth stealing.** Write the Docs posted the Tuesday lightning talks from Berlin 2026 (39:59, no chapters, so use the timestamps below). Brooke Stauner (00:00:47) reframes brevity for anyone who writes support or error documentation: your reader is already having a bad day when they open the page, so the metric isn't sentence length, it's how much work you make them do to reach the next step. If you maintain HelpHub or plugin support docs, that's the whole brief in one line. Gugulethu H. (00:17:25) adds a step most "AI router" setups skip: before picking which model, decide whether the task needs one. Year-old and small open-weight models handle most docs tasks, and pen and paper handles some. For a community that debates AI in "all or nothing" terms, "your AI" running on your own hardware is a third position worth naming. Kenneth Whyte (00:07:15) has the story every contributor who's lived through a platform migration will recognize: the site moved from GitHub to SharePoint without warning, and the Slack anger landed on the writer maintaining it, not the manager who decided. His fix was a call asking for a tour, not a justification. Also in the block: Erica Beyea on KnowledgeOwl's 25-hour week (00:23:54), Anton Kolesnyk's "paranoid positivism" and open challenge to automate your job (00:28:23), Val Grimm on the demoscene (00:33:26), and Miranda Wilson on comb jellies with a transient anus (00:12:22), which you'll want to have seen before someone brings it up at a meetup. Links: [the video](https://www.youtube.com/watch?v=dXt4A25yt2M), [Write the Docs Berlin 2026](https://www.writethedocs.org/conf/berlin/2026/). ## ✅ Action items #### DevRel - Developer advocacy: adopt Brooke's "bad day" persona (00:03:39) as the default test reader for any troubleshooting or error-reference page; measure steps-to-resolution rather than word count. - Education: run one office-hours session on G's three router questions (00:18:50 to 00:20:18), with participants classifying five of their own recurring tasks as no-AI, small-model, or frontier. - Developer experience: audit quickstart and error docs for happy-path-only coverage (00:05:49); list the three most common failure states each page ignores. - Developer marketing: no direct action from this source; nearest relevance is that a lightning-talk block is a low-cost on-ramp for first-time speakers, a pattern any community program can copy (00:00:47). - Community management: Carolyn's two-rule framing (code of conduct, time) at 00:00:04 is a reusable script for any open-mic session. - Internal/external/cross-project: Kenneth's move at 00:09:28 (call the maintainer, ask for a tour, not a justification) is a template for cross-team friction after a tooling decision. #### OSPO - Policy and governance: Kenneth's incident (00:07:58) is a case for a written "docs platform changes require notice to maintainers" rule; check whether one exists. - License compliance: G's "own the machine" stack (00:21:44) means open-weight model licenses enter the inventory; verify terms for the specific models before internal use. - Security management: a home lab running open-weight models for docs work (00:21:44) moves data off vendor clouds but onto unmanaged hardware; document the tradeoff before recommending it. - Software inventory: no direct action from this source; nearest relevance is tracking which docs tooling is self-hosted versus vendor-hosted after an insourcing push like G's (00:21:02). - Developer support: publish a one-page "do I need a model for this" decision aid based on G's router. - Community engagement: no direct action from this source. - Strategic alignment: G's aside at 00:20:18 that technical writing sits below engineering in budget hierarchy is a KPI question: what share of AI spend does docs get, and what does it produce? #### Community - Maintainers: Kenneth's "one of him" backlog observation (00:10:54) argues for a maintainer-load check before any platform migration. - Contributors: the lightning-talk on-ramp (four first-time speakers of seven) is a model for first-contribution programs; copy the low-stakes framing. - Sponsors: no direct action from this source; nearest relevance is Erica's KnowledgeOwl wage model (00:24:39) as a talking point on sustainable funding of docs labor. - Ecosystem: Anton's open challenge (00:32:01) could be run as a structured exercise: list what docs work you're paid for, then try to automate it, and publish what survived. #### WordPress - Core: no direct action from this source. - Core Performance: no direct action from this source. - Core AI: G's router (00:18:50) suggests a guardrail principle: evaluate whether a task needs a model before evaluating which model. Add it to any responsible-AI guidance. - **Core Program Team:** Kenneth's incident (00:07:58) is a coordination failure pattern; add "notify docs maintainers before platform changes" to cross-team dependency checklists. - Mobile: no direct action from this source. - Playground: no direct action from this source; nearest relevance is Val's point that hybrid, browser-reachable demos widened demoscene participation (00:37:08). - Community: the five-minute lightning format with a wrangler and two rules is a ready-made meetup segment. - Training: Brooke's cognitive-load framing (00:04:21) separates training documentation from task documentation; use it when scoping Learn WordPress lessons versus HelpHub pages. - Marketing: no direct action from this source. - Support: Brooke's persona is the support-docs reader exactly; audit the top 20 support articles for happy-path-only coverage. - Design: no direct action from this source; nearest relevance is Brooke's "unintuitive part of our UI" as a docs-generated signal for design debt (00:03:39). - Documentation: adopt "make the reader do less work to reach the next step" (00:05:06) as a style-guide principle alongside sentence-length rules; add a "do I need AI for this edit" step to any AI-assisted docs workflow. - Photos: no direct action from this source. - TV: this video has no chapters and garbled auto-captions; WordPress.tv can avoid both with speaker-timestamped descriptions and human-checked captions. - Polyglots: no direct action from this source. - Accessibility: no direct action from this source; see the accessibility assessment for what this recording gets wrong. - Meta: no direct action from this source. - CLI: no direct action from this source. - Hosting: no direct action from this source. - Tide: no direct action from this source. - Openverse: no direct action from this source. - Themes: no direct action from this source. - Plugins: no direct action from this source. #### The WPOCC - Programs: Erica's account (00:24:39) is evidence for a stipend model priced at a fair hourly rate for fewer hours rather than full-time salary; worth citing when The WP Open Community Collective designs maintainer stipends. - Projects: a documentation sprint scoped as "cover the three failure states each page ignores" applies Brooke's talk directly. - Funding experiments: no direct action from this source. ### FAIR - Technical independence: G's "own the machine" argument (00:22:27) is the same independence claim FAIR makes about infrastructure, applied to model inference; note the parallel. - Policies and covenants: Kenneth's unannounced platform move (00:07:58) is the kind of decision a notice-and-transition covenant would prevent. - Productized governance artifacts: no direct action from this source. ## 🧪 Hypotheses 1. Rewriting a troubleshooting page against Brooke's "bad day" persona (00:03:39) reduces median time-to-resolution for readers arriving from an error state by at least 20% versus the same page edited only for word count. 2. Adding G's "do I need AI for this" gate (00:18:50) before model selection cuts a docs team's paid-token volume by at least 30% with no measurable change in reviewer-rated output quality. 3. For routine docs edits (typo, terminology, restructure under 300 words), a year-old open-weight model matches a frontier model on reviewer acceptance rate within 5 percentage points, as G claims at 00:20:18. 4. When a docs platform migrates without notice, complaints in team chat name the maintainer more often than the decision-maker by at least 2 to 1 (Kenneth's observation at 00:08:43). 5. A 25-hour week at a fair hourly rate (00:24:39) produces per-hour output at or above a 40-hour comparison team's on a shared docs backlog, measured over a quarter. 6. Fewer than half of the "what keeps a technical writer employed" arguments Anton collected (00:31:17) survive a hands-on automation attempt using current skills-and-retrieval tooling. ## 📊 Methodology - **H1.** Pick 10 support articles with high entry-from-error traffic. Produce two variants each: word-count edit and persona edit (failure states covered, next step foregrounded). A/B by session; measure time from page entry to a success event (ticket closed, next page reached, or exit without return within 24 hours). Pre-register the 20% threshold. - **H2.** Instrument one docs team's AI usage for four weeks as a baseline, then four weeks with a required "no AI / small / frontier" classification before each call. Compare token spend from provider invoices and blind reviewer ratings on a sample of 50 outputs per period. - **H3.** Assemble 100 routine edit tasks from a real backlog. Run each through a named year-old open-weight model and a current frontier model with identical prompts. Two reviewers accept or reject blind; report acceptance rate and inter-rater agreement. - **H4.** Collect chat threads from five documented unannounced platform migrations (with consent). Code each complaint by target: maintainer, decision-maker, platform, or none. Report the ratio. - **H5.** Requires a willing pair of teams. Match on backlog type and seniority; track story points or pages shipped per paid hour over 13 weeks, plus a burnout instrument at start and end. A single company's self-report (Erica at 00:26:09) is the current evidence and can't settle it. - **H6.** Run Anton's unconference exercise as a structured study: collect the argument list, have two engineers attempt automation of each for a fixed two hours, and record success, partial, or fail with the artifact. Publish the list and outcomes. - **Controls.** Fix prompts and reviewers across arms; record model versions and dates; report costs from invoices rather than estimates. ## ❓ Questions - What share of a support-docs reader base arrives from an error or failure state, and does that share differ by product? Brooke's persona assumes it's the majority. - Which of G's three router questions removes the most spend in practice: not using AI at all, using a smaller model, or accepting more cleanup? - Is the "10,000 token context windows" figure at 00:21:44 what G said, or a caption error for a larger number? It changes whether the home-lab claim is plausible for long-document work. - What notice period would have prevented Kenneth's incident, and who in the organization had the authority to require it? - Does KnowledgeOwl's 25-hour week hold for roles with customer-facing SLAs, or only for work that can be time-shifted? - Which arguments did Anton collect at 00:31:17, and did anyone take the 14:15 challenge? The outcome is the interesting data and isn't in this recording. - What is the model name the captions render as "Astra" at 00:19:36 and 00:20:18? - What did the demoscene do to move from closed-source by default to increasingly open (00:37:08), and is that transferable to docs communities that hoard internal templates? ## 🔮 Future research - **Cognitive-load metrics for documentation.** Brooke's reframing needs a measurement. Time-to-next-step and re-read counts on error pages are candidate proxies; no one in this block cites an instrument. - **Model-tiering in docs workflows.** G's router is a personal practice. A published decision aid with per-task model recommendations and cost data would turn it into a team practice. - **Blame routing after platform decisions.** Kenneth's observation that maintainers absorb anger meant for deciders is testable across incident chat logs and hasn't been studied for docs teams specifically. - **Shorter-week productivity in docs roles.** KnowledgeOwl is one company. A multi-company comparison of docs output per paid hour under 25-, 32- and 40-hour weeks would settle whether Erica's experience generalizes. - **The "sensor" role.** Anton's claim that writers are paid to supply context AI can't reach is a description of a job that could be measured: how much of a writer's week is context capture versus prose production, and which half is automating faster. - **Hobbyist subcultures as learning infrastructure.** Val's demoscene account suggests that communities with explicit room for useless work retain members longer; that's an empirical question for docs communities too. ## 💡 Implications - **Practice.** Write the failure states, not just the happy path (00:05:49). Add a "does this need a model" step before any AI-assisted edit (00:18:50). After a tooling decision lands on someone else, call them and ask for a tour before asking why (00:09:28). - **Policy.** Platform changes to docs sites need a notice-and-transition rule; Kenneth's incident shows the cost of not having one. Employers pricing docs labor can price hours at a fair rate rather than headcount at a salary (00:24:39). - **Security.** G's home-lab path keeps docs data off vendor clouds but puts it on hardware without a vendor's security team; the tradeoff needs to be named, not assumed. - **Education.** Brooke's distinction between training documentation and technical writing (00:02:55) is a scoping tool for curricula: the same content serves different reader states and shouldn't share one page. Val's demoscene shows a learning community sustained by permission to be useless. - **Funding models.** G's aside that technical writing sits below engineering in budget priority (00:20:18) explains why docs teams feel token costs first. Small open-weight models are a budget strategy, not only an ethics stance. Erica's fair-hourly-wage model is a funding template for stipends and part-time maintainer roles. - **Labor.** Anton's challenge (00:32:01) is uncomfortable and useful: a profession that can't name work an automation attempt fails to reproduce needs to find that work. His answer (a second, physical role) is one option; Brooke's (writing for a reader's worst moment) and Kenneth's (peer relationships across the line) are others the same block offers without labeling them as such. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level on the transcript body: approximately grade 7 to 8. Flesch Reading Ease near 72, conversational English with short clauses. Domain terms (EHR, SharePoint, open-weight models, context windows, ctenophores, demoscene) raise effective comprehension to a general tech-adjacent adult audience, not a specialist one. #### 👀 Visual learners - Present: Miranda's slides carry the argument (comb jelly photos, a GIF of one jelly eating another at 00:13:46). Erica's before-and-after paintings at 00:26:55 are the whole evidence. Anton contrasts two pictures at 00:29:07. Val skims borrowed slides. - Improve: none of these visuals are described in speech, and the description has no slide links. Brooke, Kenneth and G have no visual scaffold at all; a one-slide diagram of G's three router questions would carry the talk. #### 🔊 Aural learners - Present: seven distinct voices and a live audience; the format is speech-first. Call-and-response at 00:37:50 and the group yell at 00:38:34. - Improve: the recording opens with people trickling in from lunch (00:11:38) and has no timestamps in the description; a per-talk chapter list would let listeners jump. #### 🗣️ Verbal learners - Present: Brooke's imperative list at 00:02:55 and Anton's coinage "paranoid positivism" at 00:28:23 are memorable phrasings. G names the trilemma (no AI, their AI, your AI). - Improve: EHR, open-weight, RAG (captioned "rack"), and demoscene terms go undefined; one clause each would cost seconds. #### 🧩 Kinesthetic learners - Present: Anton's unconference challenge (00:32:01) and Val's "Google it and come find me" are the only do-something prompts. G's home lab is a described project. - Improve: G's router could ship as a one-page worksheet; Brooke's persona could ship as a rewrite exercise with a sample error page. #### 🧠 Logical learners - Present: G's three-question decision sequence (00:18:50 to 00:20:18) is a decision tree. Brooke's argument has a clear premise-to-conclusion shape. Anton states a falsifiable claim with a deadline. - Improve: no numbers anywhere are sourced; the show of hands at 00:18:07 isn't counted. A logical learner gets structure but not evidence. #### 🧑‍🤝‍🧑 Social learners - Present: the block is explicitly social: applause for rejected submitters (00:06:32), a challenge table, a room-wide yell. Kenneth's story is about repairing a peer relationship. - Improve: no discussion prompt or follow-up channel is named for online viewers; Anton's LinkedIn post-mortem isn't linked. #### 🧘‍♀️ Solitary learners - Present: each talk is self-contained and under six minutes; a viewer can take one at a time. - Improve: no reflection prompts. "Which of your pages only covers the happy path?" and "Which of your AI calls could be pen and paper?" would convert two talks into exercises. #### 🌿 Naturalistic learners - Present: Miranda's talk is entirely natural history. Anton's second role is forest firefighting. G's argument for local models has an unstated resource-consumption angle. - Improve: G could name the energy dimension of local versus hosted inference; Miranda's invasive-species point (00:13:46) connects to ecosystem collapse but isn't developed. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies page landmarks and player controls. The description lists all seven speakers and titles in order, which is machine-readable. - Improve: slides (Miranda, Erica, Anton, Val) carry information never spoken. WCAG 2.2 AA 1.1.1 (Non-text Content) and 1.2.5 (Audio Description, Prerecorded) are unmet for those segments. Erica's two paintings are the evidence for her claim and get no verbal description beyond "this is the difference." #### Keyboard navigation - Present: the YouTube player is keyboard-operable with visible focus; 2.1.1 and 2.4.7 pass via the platform. - Improve: no chapters, so keyboard users can't jump to a talk. Adding chapters satisfies the intent of 2.4.5 (Multiple Ways) for a 40-minute multi-speaker video. #### Contrast and readability - Present: not verifiable from a transcript. Conference stage lighting and projected slides are the likely presentation. - Improve: projected slides filmed from the room typically fail 1.4.3 (Contrast Minimum) for on-screen text; insert the slide files or a high-contrast overlay in the edit. #### Text resizing and reflow - Present: the player scales and captions resize independently. - Improve: any text baked into slides doesn't reflow; 1.4.4 (Resize Text) and 1.4.10 (Reflow) fail for it. Put slide text in the description or a linked deck. #### Media and captions - Present: auto-generated captions exist; this analysis was produced from them. - Improve: the captions garble species names ("tennaphors," "neiopsis lady," "barrow or vata"), tool names ("Alama or Upfill"), a model name ("Astra"), the employer ("knowledge owl"), the conference ("Write the Docks"), "RAG" as "rack," and every demoscene proper noun. That's a 1.2.2 (Captions, Prerecorded) failure in substance. The group scream at 00:38:34 is captioned only as "[screaming]." Human-corrected captions and a downloadable transcript are the fix. #### Forms and interactions - Present: none in the video. Anton's challenge gives a time (14:15) and a location (unconference) in speech only. - Improve: not applicable to WCAG in-video. For the linked conference page, 3.3.2 (Labels or Instructions) applies to any schedule or signup form. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (inaccurate captions on every proper noun), **1.2.5** and **1.1.1** (slides and paintings conveyed visually only), **1.4.4 / 1.4.10** (slide text doesn't reflow), and **2.4.5** in spirit (no chapters for a seven-talk, 40-minute video). Chapters and corrected captions are the two cheapest, highest-impact fixes. ## 🌍 DEIB assessment #### Cultural breadth - Present: speakers reference the US military, London, Toronto, Spain, Germany, the Mediterranean, and a demoscene spanning Argentina to Japan (00:34:10). Val notes "Amiga" means friend in Spanish. - Improve: the AI-cost discussion assumes a European or North American budget context; G's "Chinese open weight models" reference (00:21:44) is the one non-Western technical reference and isn't developed. #### Inclusive language - Present: the wrangler uses she/her in the description; speakers use second person and plain terms. Anton's German employment aside (00:31:17) is stated as fact about his context, not as a generalization. - Improve: Kenneth's "war," "enemy," and Jedi/Sith framing (00:07:58) and Anton's "dead like dinosaurs" (00:29:49) are combative metaphors; both are self-aware, but plain statements would travel better across cultures. Anton's "get our [expletive] together" at 00:29:49 is bleeped in the captions. #### Neurodiversity support - Present: strong chunking by design: seven independent five-minute units with applause boundaries. - Improve: no chapters, so the chunking isn't navigable. Two speakers (Brooke, Erica) announce nerves; the format normalizes that, which helps. The group scream at 00:38:34 arrives without warning and may be a sensory problem for some viewers. #### Representation - Present: seven speakers with varied backgrounds (military instruction, contracting, marine biology hobby, home-lab technologist, visual artist, firefighter, demoscene organizer); at least four first-time speakers; a mix of genders on stage and a woman wrangling. - Improve: no speaker names or pronouns are on screen per the transcript; the description carries names but not links. Anton's talk is the only one where the audience is positioned as opponents. #### Psychological safety - Present: Carolyn names the code of conduct first (00:00:04), applauds rejected submitters (00:06:32), and frames the stage as a safe first-talk venue. Kenneth's talk is about de-escalation. - Improve: Anton's "your managers will figure out how to legally get rid of you" (00:31:17) is a threat framing that could land hard on precariously employed viewers; the invitation to argue is the safety valve and should be foregrounded in any summary. ## ✅ Logical validity #### Strong reasoning - **Causal, Brooke, 00:03:39 to 00:05:06.** Premise: the reader arrives in a failure state with high existing load. Premise: documentation adds load proportional to the work it demands. Conclusion: reduce the work per step, not only the word count. The chain is valid and yields a different editing target than "be concise." Clarify by naming a proxy metric (steps to next action). - **Structural observation, Kenneth, 00:08:43.** Complaints targeted the maintainer, not the decider. This is an inductive claim from one incident, but the mechanism (the maintainer is the visible, reachable party) is plausible and generalizes. Strengthen with a second case. - **Decision sequencing, G, 00:18:50 to 00:20:18.** Ordering the questions (need AI? need a large model? tolerate cleanup?) from cheapest to most expensive is a sound cost-minimizing structure regardless of the answers. Clarify by attaching example tasks to each branch. - **Mechanism, Erica, 00:24:39.** A bootstrap company can't afford 40 hours at a fair rate, so it buys 25 at that rate. The arithmetic holds and explains the policy's origin without appeal to ideology. The productivity claim that follows is separate and weaker. - **Falsifiable claim with a test, Anton, 00:32:01.** "Name what you're paid for; I'll automate it" is a public, time-boxed test of his thesis. Whatever the outcome, the structure is sound. Clarify by publishing the results. - **Concession against interest, Brooke, 00:05:49.** Admitting the talk restates "know your audience" raises credibility rather than lowering it. #### Fallacies - **Unsupported quantification, G, 00:18:07.** "The vast majority believe" from an uncounted show of hands. Weak because the count is a room mood, not a measurement. *Fix:* count, or say "most hands went up." - **Hasty generalization, G, 00:20:18.** "small language models are as good at tasks as they were a year ago" and "perfectly adequate for most tasks" carry no task list. *Fix:* name three tasks and one that fails. - **Anecdotal evidence, Erica, 00:26:09 and 00:26:55.** One employee's self-reported productivity and two paintings are presented as evidence the policy is good "for everyone." *Fix:* scope the claim to "for me, and here's the company's stated reason." - **False dilemma dressed as a trilemma, G, 00:18:07.** "No AI / their AI / your AI" omits the mixed case most teams live in (hosted models under a data agreement). Low severity; the rhetorical move is to add an option, which is the opposite of narrowing. *Fix:* name the hybrid. - **Appeal to fear and unsupported deadline, Anton, 00:31:17.** "Maybe till new year" and "legally get rid of you" carry no basis. Anton signals it's a gut estimate and invites challenge, which mitigates but doesn't cure it. *Fix:* state the evidence for the horizon or drop the date. - **Slippery slope, Anton, 00:29:49.** Dinosaurs and a meteor imply extinction from current trends without intermediate steps. *Fix:* name the specific tasks that have already been automated and the ones next. - **Unsourced natural-history claims, Miranda, 00:14:29 to 00:16:42.** Regeneration from a third, fusion, reverse aging, hourly transient anus. Plausible and consistent with popular accounts but no paper is named. *Fix:* one citation per claim in the slide footer. Not a fallacy so much as a sourcing gap. - **Caption-induced ambiguity, throughout.** Not the speakers' fault, but "Astra," "10,000 token," "Alama or Upfill," and the demoscene names can't be cited as fact from this transcript. *Fix:* corrected captions. #### Tone of voice - Warm, informal, self-deprecating, and fast, in the standard lightning-talk register. Brooke jokes about her own verbosity while being verbose; Kenneth uses Star Wars framing to lower the stakes of a real grievance; Erica apologizes for nerves and says "I studied art, whatever" while showing gallery-grade work. G is the most earnest and direct ("shady companies that have stolen everyone's data," 00:22:27). Anton is deliberately provocative and says so ("confident but paranoid king," per Carolyn at 00:32:44), and the provocation is bounded by an explicit invitation to argue. Val is the most performative, ending in a group yell. The wrangler's tone (two rules, applause for the rejected, "I could do that") sets a floor of safety the provocative talks lean on. The most useful reading note: the block's emotional range is the point; the content varies, the permission to try is constant. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response to accusation. No speaker is accused of anything, and no one responds to criticism defensively. The nearest material: Kenneth's story is about being on the receiving end of blame for a decision he didn't make (00:08:43), and his response is the opposite of DARVO, a de-escalating call to the other party. Anton pre-empts disagreement by inviting it (00:32:01) rather than attacking critics. For completeness across the three prongs: **Deny**: no denials; Brooke and Erica volunteer their own limitations (verbosity, nerves, self-report). **Attack**: G criticizes AI vendors generically at 00:22:27 ("shady companies"), which is a stance, not an attack on an accuser; Anton's "dead like dinosaurs" targets a discourse, not a person. **Reverse victim/offender**: no such move. Kenneth describes the writer as the one hit by the Slack war, which is a factual account of blame routing, not a reversal. ## 🎯 Conclusion Three of these seven talks give a documentation strategist something to use this week. Brooke's argument at 00:03:39 to 00:05:06 converts "be concise" into a measurable target: the reader is already failing when they arrive, so the page's job is to reduce the work between them and the next step, including on the unhappy paths. G's router at 00:18:50 adds the step most AI workflows skip, deciding whether a model is needed before deciding which one, and pairs it with a budget reality (docs teams feel token costs first). Kenneth's observation at 00:08:43, that anger about platform decisions lands on the maintainer rather than the decider, comes with a practiced response: call the person on the other side and ask for a tour, not a reason. The rest is context. Erica's 25-hour week is one company's fair-hourly-wage policy with one employee's testimony; Anton's "till new year" horizon is a gut estimate attached to a falsifiable challenge whose outcome isn't in the recording; Miranda's comb jellies and Val's demoscene are the block's argument, never stated, that a docs community needs work that is useless on purpose. **Next steps for readers:** pick your top 10 support or error pages and mark which ones cover only the happy path; that list is the Brooke backlog. Log a week of your own AI calls and classify each as no-model, small-model, or frontier; the first column is the savings G is describing. If you're mid-migration or just past one, find the person maintaining the new platform and ask for a tour. And if you cite any of the numbers or names from this recording, check them against the speakers' own materials first, because the auto-captions got most of the proper nouns wrong. ## 🔄 Refactor Appendix — Tuesday Lightning Talks - Write the Docs Berlin 2026 > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied: Miranda's and Val's > talks were dropped from the atoms as enjoyable but not transferable to docs practice; > Anton's "till new year" horizon was dropped as an unsourced deadline. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Docs readers arrive already having a bad day > [!quote] [00:03:39] > the person reading my documentation is doing so because they're having a really really bad day. **Context & use**: - 00:03:39, Brooke Stauner, after listing brevity imperatives and admitting knowing them isn't practicing them. - This is the persona that makes "know your audience" operational for support and error documentation. The reader's existing load is the constraint; the page's job is to add none. **Tags**: #xref #aphorisms #share <!-- atomicity: 1 idea ✓ --> #### [H2] Tooling wars hit the maintainer, not the decider > [!quote] [00:08:43] > They were hitting the writer who was having to update and maintain the site. **Context & use**: - 00:08:43, Kenneth Whyte, describing the Slack fallout after a docs site moved from GitHub to SharePoint without notice. - The mechanism: the maintainer is the visible, reachable party. Any cross-team escalation process should route grievances to the decision owner explicitly, or the maintainer eats them by default. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H3] Decide whether to use AI before choosing a model > [!quote] [00:18:50] > I find that we miss the step before that when you have to decide if you want to use AI in the first place. **Context & use**: - 00:18:50, Gugulethu H., introducing a personal router whose first branch is "no model," with pen and paper as the stated alternative. - Most model-routing setups start at model selection. Adding the prior gate is a cost lever and a craft lever at once, and it's the part of the talk that applies to any team regardless of which models they run. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H4] A shorter week can price labor at a fair hourly rate > [!quote] [00:24:39] > So rather than trying to get someone to do more hours for the same amount of money, it's a fair wage. **Context & use**: - 00:24:39, Erica Beyea, explaining why KnowledgeOwl started a 25-hour week as a bootstrapped company. - The origin is an affordability mechanism, not a wellness program: buy fewer hours at the rate people are worth. That framing transfers to stipends and part-time maintainer roles in a way "four-day week" advocacy doesn't. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### [H5] Writers currently earn pay as AI's tactile sensor > [!quote] [00:30:32] > So it has a finger it's a tactile sensor and it just touches the stuff that it cannot touch without me **Context & use**: - 00:30:32, Anton Kolesnyk, describing what he's paid for "for now" in a talk he calls a call for help. - Uncomfortable and precise: the current human contribution is context capture, not prose. Useful as a job description to test against, and as a prompt for which half of the work is automating faster. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> ### 🧠 Atomic Split Pad #### [P1] Write for the reader's worst moment **Claim** (one sentence, Courtney's voice): Support documentation earns its place by reducing the work a reader who is already failing must do to reach the next step, not by completing the writer's picture of the system. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds for troubleshooting, error reference, and task documentation reached from a failure state. Does not hold for conceptual or training material, where the reader chose to arrive with time and attention; Brooke's own distinction at 00:02:55 marks the boundary. Breaks if "less work" is read as "fewer words" and the unhappy paths get cut. **Links**: - [[Cognitive load in documentation]] — because extends - [[Diátaxis quadrants]] — because applies-in separating how-to and reference from explanation and tutorial - [[Every claim carries a fact]] — because contrasts on target: reader effort versus writer precision <!-- atomicity: 1 idea ✓ --> #### [P2] Gate the model call before routing it **Claim** (one sentence, Courtney's voice): An AI workflow that first asks whether the task needs a model at all saves more than one that only picks between models. **Evidence pointer**: [[#H3]] **Scope & conditions**: Holds for individual and team docs work where tasks vary in size and many are small; the gate is cheapest when the human already knows the answer or the task is under a few hundred words. Weakens when tasks are uniform and always need generation, where the gate becomes overhead. The budget aside at 00:20:18 (docs below engineering in spend priority) is the condition that makes the gate matter most. **Links**: - [[Classify before you generate]] — because extends the cost-tiering idea one step earlier in the pipeline - [[Credit routing]] — because applies-in deciding when to delegate at all versus which model to delegate to - [[Local model self-hosting]] — because evidence-for the "your AI" branch G proposes <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Recover the model name captioned "Astra" at 00:19:36 and 00:20:18, and the tool captioned "Upfill" at 00:21:44, from G's own materials before citing the talk: https://www.youtube.com/watch?v=dXt4A25yt2M&t=1176 #fleeting #inbox - [F2] Check whether Anton posted the surviving "what keeps writers employed" list on LinkedIn after the 14:15 unconference table; the outcome of the challenge is the interesting data and isn't in the recording #fleeting #inbox - [F3] Open question: what share of HelpHub or plugin support page traffic arrives from an error or failure state? That number sizes the Brooke backlog #fleeting #inbox - [F4] Article seed: "The router step everyone skips," G's no-model gate as the missing first branch in AI cost tiering, paired with the classify-before-generate note #fleeting #inbox #articleseed - [F5] Verify the ctenophore claims (regeneration from a third, fusion, reverse aging, hourly transient anus in *Mnemiopsis leidyi*) against primary literature before repeating them; the talk names no source #fleeting #inbox - [F6] Kenneth's next-year teaser: a CTO publishing AI-generated docs from a personal GitHub repo. Watch for that talk at Write the Docs 2027 #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Mark the top 10 support or error pages I own as happy-path-only or failure-state-covered — docs audit, 30 min #things - [ ] Log one week of my own AI calls and classify each as no-model, small-model, or frontier — AI cost review #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Tuesday Lightning Talks - Write the Docs Berlin 2026.md|Tuesday Lightning Talks - Write the Docs Berlin 2026 by Write the Docs]] ## Highlights - [00:02:55] documentation needs to be on target. Get to the point. Say more or less. Be concise. Stop conflating training documentation with technical writing. High speed, low drag. Embrace brevity. But how do I get there? And how do I stay on that path? So, we already talked about this a lot here. Use your tools, right? Your AI tools, your style guides, use your community, uh, communicate and peer review. Um, but there's knowing that and then there's practicing it and staying consistent with that, especially for a new writer. Um, for me, I need help starting at the point of brevity. And I think that, uh, the amount of rework that I do to whittle down the context that I want my users to have to get to the core of what they need, uh, [00:03:39] it's a lot of rework. And I think about what they need at the point of need, uh, because of course I'm thinking about my audience. And so I do that by remembering this. Uh the person reading my documentation is doing so because they're having a really really bad day. That person who needs my documentation, that emergency medicine doctor or that medic who's totally disconnected. Um they're all the way forward and as lovely as I write it, they're not looking at my documentation because they want to uh optimize their workflow. Um they're getting an error message. They're stuck on an unintuitive part of our UI. They're trying to give a patient a medication and communicate to the nurse without doing harm. Um, and they're trying to do it quickly and ([View Highlight](https://read.readwise.io/read/01m37b7nb9m3e548rxt17e65qt)) - Note: [H1] Docs readers arrive already having a bad day - Tags: atom/literature/h1 - [00:08:43] who um were managing the site. So um the developers and my management dealt with this in the um agreed manner and went to war on the Slack channels. Um, and what I noticed about um, the Slack war was that they weren't really hitting the manager whose decision it had been, and they weren't really hitting the the senior staff whose responsibility was creating and managing the site. They were hitting the writer who was having to update and maintain the site. So someone like me and for all that um we could sit and argue about the merits of um GitHub over SharePoint um I [00:09:28] still had to find a way of displaying the information that I needed to on the page I was working on. So, I'd met the writer briefly before and I messaged him and arranged a call and rather than ask him about why it happened or or you know what was the reasoning behind it, I just asked him to take me through the site. Um there were certain areas that um I you know needed to clarify where I could find you know information points that we needed to share. Um and I also just talked to him about his way of working. you know, did moving the site like this mean, you know, did it make it easier for him in certain ways more difficult? And he went from being quite anxious from talking to someone in ([View Highlight](https://read.readwise.io/read/01m37b7qhz1mw5brdfhk9p6bmk)) - Note: [H2] Tooling wars hit the maintainer, not the decider - Tags: atom/literature/h2 - [00:18:50] the frontier themselves. And I feel like a lot of the conversation that we have is oified around I hate AI or you're going to get left behind. And um there's a third option there which I feel like we we don't talk about enough. And something that I've been doing recently um just to help me to to think about situations and where and when I use AI is a personal router. Like we all know a router for AI where you can decide that um I'd rather use Claude for code or uh chat GPT for um general use. I find that we miss the step before that when you have to decide if you want to use AI in the first place. Um, and one of the things that I've missed as a [00:19:36] someone who loves writing is pen and paper. And there are times when I feel like craft is uh a really enjoyable thing. Um, and actually writing something out feels like it sorts things out in my head. And this personal routter process for me really helps me decide whether I need to use AI, whether I need to use a frontier model. Um, I know some of you maybe have already started testing Astra and seeing how well it works and also how expensive it can get and very very quickly. Um, and the reason um, uh, I ask myself this question is because most of the time things that came out a year ago or two years ago are perfectly adequate for ([View Highlight](https://read.readwise.io/read/01m37b7stfa4jk3y8m1v3h0njd)) - Note: [H3] Decide whether to use AI before choosing a model - Tags: atom/literature/h3 - [00:24:39] sort of thing like finding your way to it. Similarly, Knowledge started the 25-h hour work week because they're a small bootstrap company and that's what they could afford when they wanted to pay people um the amount that they're worth per hour. So rather than trying to get someone to do more hours for the same amount of money, it's a fair wage. And yeah, anyway, so that's sort of how that started. But here to talk about my experience with this is I'm a visual artist before working in um software at all. I was a bartender, an art teacher, a painter. I worked I lived in Toronto, expensive rent. I worked like five part-time jobs [00:25:23] and painted. [snorts] Um, and I found myself regularly either having time to paint and no money to eat or um, money to eat but zero time to paint and uh, accidentally ended up in software because I moved to Spain without any savings or a plan. Um, and needed a job and found a remote job and uh, so but there I was still working like 40 plus hours a week. I ended up managing the CX team over time and I had again very little time to paint and I came across a knowledge job post and I was like 25 this is insane like I couldn't believe it was real. Um and I ended up applying ([View Highlight](https://read.readwise.io/read/01m37b7w3kw37eab96d6sdzxk5)) - Note: [H4] A shorter week can price labor at a fair hourly rate - Tags: atom/literature/h3 - [00:30:32] oh I'm a technical writer. I used to be a technical right. I'm a developer. I'm a content designer blah blah blah many many professions. I am a sensor. Yeah. And I do feel like this because what I feel now I'm doing at work AI is not yet capable enough to grasp the whole world context and it needs me to I don't know I that's how I see it. So it has a finger it's a tactile sensor and it just touches the stuff that it cannot touch without me and then that's what I'm getting paid for for now. Yeah. So um and uh based on the conversations that I had this um day and then previous days on this beautiful conference, thank you very much for everyone for contributing um and I came up with a set of uh points that people think will keep [00:31:17] their job running for the next couple of months uh by my estimates. Um so but I don't still believe that any of this hold brown because if you look at it uh critically and then if you're familiar with how I now works and you know how to do skills this rack stuff all these beautiful things you can copy past from YouTube and everybody everything else it doesn't hold the ground maybe till new year before your managers if you're in Germany based they will figure out how to legally get rid of you uh but I don't think it will hold so therefore I'm invite you inviting asking all of you or any of you who are ready to argue with me and challenge my um idea that nothing will hold you here long labor at uh ([View Highlight](https://read.readwise.io/read/01m37b7yf36yxe30k23f36jv7v)) - Note: [H5] Writers currently earn pay as AI's tactile sensor - Tags: atom/literature/h3