Episode 36: Jacquie Capur  🔗 **Source:** [Episode 36: Jacquie Capur](https://www.youtube.com/watch?v=V9u-iN1tx4I) > [!note]+ Document Notes > tags: [devrel, developer-community, ai-assisted-development, devrel-metrics, live-coding] > author: Jacquie Capur (guest; the transcript captions spell the name "Jackie Kapor"), Chris Reynolds (host) > source: Community & Code (YouTube) > reference: https://www.youtube.com/watch?v=V9u-iN1tx4I > date created: 2026-10-17 19:51 > > > [!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 Do not quote or attribute any other line to the named author.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: Episode 36: Jacquie Capur > <!-- 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--episode-36-jacquie-capur) ## 📝 Summary - **What happened.** The Community & Code channel published episode 36, a 43:28 interview uploaded 2026-10-17. Host Chris Reynolds talks with Jacquie Capur, a DevRel practitioner who has worked at AWS, HashiCorp and (currently) Nebius, per the guest's own account at 00:01:39. The video has no chapters, and the transcript is auto-captioned. - **Name flag.** The transcript captions spell the guest "Jackie Kapor" at 00:00:13 and "Jackie Kapor on LinkedIn" at 00:42:45. The video title says "Capur," and the LinkedIn URL in the description ends in `jacquie-capur`. Two description links (Sessionize and the HashiCorp author page) use `jacquie-grindrod`. The transcript never says "Grindrod," and its sound-alike is "Kapor," so this note uses "Capur." The transcript does not explain the second surname; confirm it with the guest or the episode page before using it. - **The claim.** The guest argues that AI tools remove production toil and bring back the "wonder" of early-career tech (00:21:02 to 00:23:12), while a human still has to translate engineering "geek speak" and supply taste (00:17:26). The guest also reports that DevRel measurement has moved from GitHub stars and signups to second- and third-layer signals such as non-employee answers in a Discord (00:31:09). - **Why it matters.** The episode gives a practitioner's first-person account of three live DevRel questions: how to measure the work, how to use agents without losing a personal voice, and what changes when content volume floods every channel. It is one person's experience, not a study. - **Key stakeholders.** The guest (DevRel at Nebius); the host (manages a DevRel team, per 00:30:25); CodeTV and the cast and crew of its live-coding competition show; DevRel teams under pressure to show business impact; developers receiving AI-written content. - **Risks and tradeoffs.** Every figure in the episode is self-reported and unsourced: the 142 VMs, the 16-agent pipeline, the 450 Zoom sign-ups, the TikTok-versus-Google search claim, and the host's Gen Z survey. The intro gives the audience at a speaking milestone as 1100 people at DevRelCon, while the guest's own account at 00:04:33 names a DevOps Toronto main stage and the caption drops the number. Several proper nouns are caption garbles. - **Bottom line.** Useful as a primary account of how one DevRel practitioner works with agents and thinks about metrics. Treat its trend claims (content saturation, AI realism, open-weights resurgence) as predictions from experience, and verify the names before citing. ## 📖 Main content **Format and framing.** This is an interview podcast. The host sets the frame in the cold open: public speaking fear, a music-major background, a live-coding reality show, and "wonder" (00:00:13 to 00:01:39). The guest answers; the host adds opinions and anecdotes of their own at 00:14:29, 00:23:56, 00:30:25 and 00:36:11. This note labels each as **Guest** or **Host**. **1. Career path (guest, 00:02:22 to 00:09:33).** The guest studied music, built Neopets and MySpace profile pages as a teenager in the early 2000s, asked a parent whether "those computer things" could be a career, switched to computer science, and took a DevOps role after being offered a choice between two job descriptions. The pull was the "full picture": tracing where data comes from and how it reaches a pipeline. The early work involved backup tapes and copying files to servers by hand. **2. Speaking as exposure therapy (guest, 00:03:49 to 00:06:43).** The guest froze in meetings that went beyond the immediate team plus a couple of stakeholders, and spoke at engineering all-hands of roughly 300 people. The fix was a bigger stage: one meetup talk, then a main-stage talk. The guest kept a gifted stress stone as a good-luck charm and carried it to a TV finale. The host notes that several DevRel guests share this pattern of speaking to get past a block. **3. The CodeTV show (guest, 00:09:33 to 00:13:04).** A friend suggested applying to a CodeTV live-coding series. The guest applied with their husband as the partner, went through weekly eliminations among eight teams, and flew to Portland every week for filming. The guest describes it as "Great British Bake Off" with coding. The description links the series as *The Full Stack* (season 1); the transcript never names it. **4. Launch Buddy (guest, 00:13:04 to 00:14:29).** The team built an agent product called Launch Buddy for DevRel launch stories. The guest reports a 16-sub-agent pipeline that compares copy to other posts, proposes 20 hooks, tests them, runs evals and returns a template. A judge then asked for a version that takes a GitHub link instead of a form. The guest deliberately stopped the tool before generation: "I still want to be the person on my videos." **5. AI-written text and the human layer (host and guest, 00:14:29 to 00:18:10).** The host describes reviewing a pull request and a marketing blog post that read as machine-written and were confusing; a colleague confirmed the blog post was AI-written. The host states at 00:17:26 that DevRel jobs are not at risk "yet." The guest adds a human translation layer, and the host adds a "taste layer," as what readers trust. Speaker attribution for 00:15:12 to 00:17:26 is inferred: the caption's turn markers are unreliable, and the host's role as DevRel manager (00:30:25) and the marketing-review story point to the host. **6. Toil, whimsy and demos (guest, 00:18:10 to 00:21:02).** The guest's "happy path" is to automate toil so there is room for whimsical projects. Examples: terminal visuals, a St. Patrick's Day demo, a take-a-photo-of-your-dog game with sprite sheets built from text-to-text models judged by a second model in an eval loop, and a short-video workflow that cuts dead space and adds lower thirds, title cards and animations. **7. Wonder and a lower barrier to building (host and guest, 00:20:18 to 00:25:20).** Both describe a lower cost from idea to working prototype. The guest describes early DevOps as walking through a castle where each door held something new: configuration management with Salt after manually updating about 142 VMs, then git, then pipelines. The host asks whether the guest feels less rushed; the host feels pressure to output more with the same resources. The guest answers that the pressure dates to about 2020 to 2021 and that this is the first time the tools feel like they sit in the guest's own hands. **8. Saturation and the human voice (guest, 00:25:20 to 00:26:45).** The guest predicts that audiences will weigh authenticity more heavily as feeds fill with obviously scaffolded posts. **9. DevRel 2020 versus now (guest, 00:26:45 to 00:29:38).** The guest started in DevRel in 2020 and describes the shift from face-to-face conferences to live streams, podcasts and Zoom workshops during the pandemic (a 400-person Zoom limit against 450 sign-ups), then to a go-to-market (GTM) emphasis with more pressure to attribute business outcomes. The guest sees the role "maturing" and compares DevRel measurement to observability. **10. Measuring DevRel (host and guest, 00:29:38 to 00:35:28).** The host says much of DevRel is "vibes" and had Claude build a measurement dashboard. The guest says three years ago the answer was GitHub stars and signups; now the guest looks at non-employee answers in a Discord and whether insights help users stick, adopt features and help others, which the guest calls a "pyramid scheme of helping." The guest tracks posts and meeting notes through an agent workflow into a dashboard, and shifted focus toward short video after the claim that TikTok overtook Google search. The guest names answer engine optimization (AEO) as a thing to track: "Can robots read our docs?" **11. AI realism and open weights (host and guest, 00:35:28 to 00:39:04).** The host reports going back to plain web search for tasks that do not need an agent, cites an unnamed survey saying Gen Z is both the most critical of AI and its heaviest user, and argues critics need hands-on use. The guest predicts price increases, comparing AI pricing to early Uber and DoorDash, and a resurgence of open-weights models because proprietary models change with short notice. **12. Whimsy and hobbies (host and guest, 00:39:04 to 00:43:28).** The guest returns to music: using Strudel with Claude or Codex while keeping the harmony and bass-line decisions personal, then writing and producing an anniversary song for their husband on a MIDI keyboard without Strudel. The guest argues technologists should not need to give every waking hour to the work. Contact handles are given at 00:42:01 and 00:42:45. **Unsourced or unverifiable figures in this source.** 1100 strangers at DevRelCon (host intro, 00:00:13); a main-stage audience whose number the caption lost (00:04:33); 300-person company all-hands (00:03:49); eight teams and six weeks (00:10:15 to 00:11:39); 16 sub-agents and 20 hooks (00:13:47); about 142 VMs (00:22:28); 450 sign-ups against a 400 cap (00:28:10); TikTok "getting more searches than Google" (00:32:37, a widely repeated claim with no source here); the Gen Z survey (00:36:53); AI pricing "already" rising (00:38:21). **Caption garbles and likely forms.** "Devril" is DevRel; "Hashi Corp" is HashiCorp; "Nebus" is Nebius (the description confirms the spelling); "Devrocon" and "Devrilcon" are DevRelCon; "Jason Langdorf" is likely Jason Lengstorf (CodeTV); "Kelsey High Tower" is likely Kelsey Hightower; "Darko Msos from AWS" is unverified (possibly Darko Mesaros); "cloud code" at 00:20:18 is Claude Code; "CLMs" at 00:25:20 is probably LLMs; "spinner bus" is probably Sprinter bus; "air program" at 00:05:58 is probably pair program; "scance" at 00:41:18 is probably askance; "Kimmy aspect" at 00:28:53 is unclear; "Magic Girl keyboards" and "melodics" switches at 00:42:01 are unclear; "non-emp employees" at 00:31:09 is probably non-employees. Quote blocks below keep the caption spelling. **Timeline oddity.** At 00:05:58 the guest places the stress stone and the TV finale around "2018 2018 2019," yet the guest says the DevRel career began in 2020 and the show involved building agents. The sentence is ambiguous: the date may refer to when the stone was gifted, not to the finale. ## 🗒️ Notes #### ⏱️ [00:00:13](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=13) Cold open and guest background - Host: previews the episode: speaking fear, a music-major start, the reality show, "wonder," what DevRel measures, the human translator, an anniversary song. - Host: names the guest's employers as AWS, HashiCorp and Nebius; the guest confirms AWS, HashiCorp and Nebius at 00:01:39 and adds five-plus years in DevRel. - Guest: studied music, then computer science; career background is platform engineering and DevOps (infrastructure as code, orchestration, pipelines). - Guest: built Neopets and MySpace profile pages with HTML tables in the early 2000s and did not see it as a career until a conversation with a parent. > [!quote] [00:00:13] > Speaking in front of 1100 strangers at Devrocon. > [!quote] [00:03:05] > do you know like those computer things that I do? #### ⏱️ [00:03:49](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=229) Fear of speaking leads to DevRel - Guest: dreaded meetings beyond the immediate team and a couple of stakeholders; the company-wide all-hands (300-something people) wrecked the morning before and needed an hour to recover. - Guest: chose a scarier setting to get past it: one meetup talk, then a main-stage talk at what the caption renders "DevOps Toronto." The audience number is lost in the caption ("like,00 people"). - Host: the intro says 1100 people at DevRelCon. The transcript supports one large stage the guest names as DevOps Toronto, and does not connect it to DevRelCon. - Host: asks why so many DevRel people start introverted yet take a public role (00:05:16); the guest says the guest was always the chatty engineer. - Guest: a teammate gave the guest a stress stone, later packed as a good-luck charm for a TV finale. > [!quote] [00:03:49] > I would like vibrate like I'd get so nervous > [!quote] [00:06:43] > yourself, challenge yourself, see where you can grow and keep the people that support you close. #### ⏱️ [00:06:43](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=403) Early computing and the move into DevOps - Guest: the guest's father, a school vice principal, played Wolfenstein with the guest at about age four after finishing evening paperwork, which made computers feel normal; the guest also had a GameCube and N64. - Guest: after about a year of computer science, a friend of a friend offered two job descriptions without labeling them; the guest picked the one that offered the "full picture" of data and pipelines. - Guest: early tasks were backup tapes and copying files to servers by hand; the guest enjoyed making things "a little more efficient." > [!quote] [00:07:25] > do you want to come sit with me and we'll play Wolfenstein. #### ⏱️ [00:09:33](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=573) The CodeTV live-coding show - Guest: the show ran through CodeTV with a person the caption renders "Jason Langdorf" and a host the caption renders "Kelsey High Tower"; both are likely garbles, and the transcript does not state Jason's role (see Main content). - Guest: applied with their husband as the partner; the application asked who the partner was and why they would fit. The guest had live-coded on Twitch for years and had never built a startup or product. - Guest: eight teams, one eliminated each week, a no-code pitch week first; the cast shared meals four days a week for up to six weeks. - Guest: flew to Portland weekly; 12-hour work days around filming; two roughly 12-hour studio days with about two four-hour filming blocks. > [!quote] [00:10:15] > it's like Great British Bake Off, but you're coding and building an app instead of building a cake or pastry. > [!quote] [00:11:39] > if you'd asked me if I could have built this in one day, I would have told you no. #### ⏱️ [00:13:04](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=784) Launch Buddy and the human layer - Guest: the guest and husband (an engineering manager for an SRE team at a security company, per the guest) wanted to try agents; the guest's DevRel pain point was building a launch story. - Guest: Launch Buddy takes an idea and returns scripts, hook candidates and a messaging template; a judge pushed them to replace the intake form with a GitHub-link input. - Guest: stopped the tool short of generating the finished piece to keep voice and story human. - Host (attribution inferred from caption turn markers): describes a pull request (code fine, description long and machine-like) and a confusing marketing blog post that a colleague confirmed was AI-written; says the project requires disclosure of AI assistance (00:15:12). - Guest: positions the DevRel skill as translating engineering "geek speak" into plain English (00:17:26). - Host: adds a "taste layer" that goes missing when an agent does the writing alone (00:18:10); says, as opinion and not finding, that DevRel jobs are not at risk yet. > [!quote] [00:13:47] > I think it was like a 16 sub agent orchestration pipeline > [!quote] [00:14:29] > I still want to be the person on my videos. I still want it to be my story and my voice > [!quote] [00:17:26] > I don't feel like we are at risk of losing our jobs to to AI yet. #### ⏱️ [00:18:10](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=1090) Toil, whimsy and demos - Guest: "happy path" is DevOps-style toil removal that frees time for better ideas. - Guest: demos the guest could not previously build: terminal visuals, a St. Patrick's Day piece, a dog-photo-to-video-game project in which text-to-text models draw SVG sprites and a second model judges them in evaluation loops. - Guest: a short-video workflow now cuts dead space and adds lower thirds, title cards and animations; editing time had been the blocker to filming. - Guest: calls the current moment like the early-2000s internet, "new and shiny again." > [!quote] [00:18:10] > my happy path is like it's very similar to DevOps like automate away the toil. > [!quote] [00:20:18] > it feels like this era of the early 2000s internet where everything's kind of new and shiny again #### ⏱️ [00:21:02](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=1262) Wonder, the lower barrier and the pressure to do more - Host: says Claude Code shrank the gap from a weird idea to a working prototype, which revived side projects; recalls learning on GeoCities by viewing source and copying. - Guest: agrees "wonder" is the right word; describes the early DevOps career as a castle of doors: Salt configuration management after updating about 142 VMs by hand, then git, then pipelines. - Guest: the sense of novelty faded as the volume of tools grew; the guest wanted either to recover the joy or to pause and feel proud of what had been built. - Host: feels pushed to output more with the same resources and says that is common across tech. - Guest: the "do more with less" pressure dates to about 2020 to 2021; earlier answers were logistics and business processes, and this time the guest has tools that take chunks out of the work. > [!quote] [00:22:28] > I open a door and I'm like oh configuration management. > [!quote] [00:23:56] > I feel like I'm I'm being asked to output more with the same amount of resources. > [!quote] [00:24:37] > the pressure is still there but it's the first time I feel like I have like shiny new tools that are at my disposal #### ⏱️ [00:25:20](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=1520) Content saturation and DevRel then versus now - Guest: "everything's a pipeline" is the guest's DevOps mindset; says there is pressure to post daily on every platform, yet audiences are oversaturated and starting to pull back from obviously scaffolded posts. - Guest: predicts audiences will weight insight, thoughts, feelings and opinions more heavily again. - Host: loses a thought mid-answer at 00:26:02 and moves on; the guest answers "it happens." - Guest: entered DevRel in 2020, when everything went online; had streamed World of Warcraft on Justin TV since about 2011, so live streaming was already in the guest's kit. - Guest: the central 2020 problem was building rapport with people who are not in the room; the pandemic shift went from conferences and meetups to live streams, podcasts and Zoom workshops. - Guest: a 400-person Zoom limit against 450 sign-ups forced a scale-out question. > [!quote] [00:25:20] > There's too many blog posts every day. There's too much content every day. > [!quote] [00:26:02] > I think those are things that are going to start getting weighted more heavily again. > [!quote] [00:27:28] > how do we make online spaces feel like real life which is obviously very difficult and challenging. #### ⏱️ [00:28:53](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=1733) DevRel measurement and the GTM pull - Guest: DevRel is trending toward GTM, with more effort to attribute business outcomes to a podcast or a talk; the guest says this pulls people toward being "content factories" and away from relationship work, and that the role is maturing, not declining. - Host: since becoming manager of a DevRel team, feels a need for numbers; says much of DevRel is vibes; had Claude build a measurement dashboard. - Guest: three years ago the answers were GitHub stars and signups; now second- and third-layer signals count, for example non-employee answers in a community Discord, whether insight helps users stick and adopt features, and whether it helps them help others. - Guest: "pyramid scheme of helping" is the guest's description of scaling beyond one-to-one help; jokes it will be quoted out of context. - Guest: admits to drawing a blank on favorite metrics during the interview (00:31:09). > [!quote] [00:28:53] > we're trying to find more ways to measure Devril > [!quote] [00:29:38] > kind of like observability but for devro > [!quote] [00:30:25] > much of Devril is vibes > [!quote] [00:31:09] > it's not just how many people joined our community, it's how many non-emp employees answered questions in the Discord, right? > [!quote] [00:31:53] > we help these two people, they help two people. Pyramid scheme of helping. #### ⏱️ [00:32:37](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=1957) Agent dashboards, platform shifts and AEO - Guest: questions whether Twitch is still the right platform as audience share moves from millennials to Gen Z. - Guest: cites TikTok overtaking Google for searches as a reason metrics now center on short video; no source is given. - Guest: runs an agent workflow over meeting notes (topics, deadlines, demo ideas) and over post metrics into a dashboard, to see growth, flat trend or decline; the same post performs differently on LinkedIn and X. - Host: says the host had no answer either until asking Claude; calls the target a moving one. - Guest: names AEO as unavoidable: people put search questions into agents, so docs must be readable by robots as well as people. > [!quote] [00:32:37] > Tik Tok was getting more searches than Google > [!quote] [00:35:28] > Can robots read our docs? #### ⏱️ [00:35:28](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=2128) AI realism, Gen Z use and open weights - Host: the "AI will solve everything" idea has pulled back and that is healthy; the host went back to plain search for tasks that do not need an agent. - Host: cites an unnamed recent survey that Gen Z is the most critical of AI and its heaviest user, and argues critics must use the thing to judge policy; makes a jab at lawmakers with flawed assumptions. - Guest: agrees feedback needs use; predicts prices will rise as with early Uber and DoorDash, and a resurgence of open-weights models because proprietary tools can change with 24-hour price cycles that used to take months or years. - Guest: says the guest has no strong view on policy. > [!quote] [00:36:11] > I've actually switched back to doing the actual search now > [!quote] [00:36:53] > Gen Z is is like the most critical of AI, but they also are the demographic that are using it the most. > [!quote] [00:38:21] > already seeing the price starting to go up. > [!quote] [00:38:21] > I do think we're going to see a resurgence of open source. #### ⏱️ [00:39:04](https://www.youtube.com/watch?v=V9u-iN1tx4I&t=2344) Whimsy, music and the close - Host: asks what adds whimsy to a bureaucratic tech life. - Guest: rediscovering a music background with Strudel and Claude or Codex, keeping the guest's own decisions on walking bass and chords instead of letting a model generate the music. - Guest: wrote and produced an anniversary song for their husband on a MIDI keyboard, playing every chord, without Strudel. - Guest: early technologists were hobbyists; people should not have to give every waking hour to count as technologists; lists a stream countdown, music theory and composition, keyboards and about 60 plants as small joys. - Close (00:42:01 to 00:43:28): the guest's handle is "DevOps Jackie" on Twitter, Bluesky and elsewhere; "Jackie Kapor" (caption) on LinkedIn; Instagram is returning. The host thanks the guest and the audience. Music plays under the sign-off. > [!quote] [00:40:34] > wrote and produced a song for my husband > [!quote] [00:41:18] > I don't think that we should necessarily have to dedicate every waking hour to be considered a technologist, right? > [!quote] [00:42:45] > I'm bringing back Instagram. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. - [Episode page: Episode 36: Jacquie Capur: Community + Code](https://communitycode.dev/episodes/episode-36-jacquie-capur/): from the description; the canonical episode write-up. - [Community + Code](https://communitycode.dev): the podcast's site. - [Jacquie on Sessionize](https://sessionize.com/jacquie-grindrod): speaker profile; uses the `jacquie-grindrod` slug. - [HashiCorp author page](https://hashicorp.com/en/blog/authors/jacquie-grindrod): also uses the `grindrod` slug. Together with the LinkedIn slug below, this is why the surname needs confirming. - [Jacquie on LinkedIn](https://www.linkedin.com/in/jacquie-capur/): spoken at 00:42:45 as "Jackie Kapor on LinkedIn." - [@devopsjacquie on X](https://x.com/devopsjacquie) and [@devopsjacquie on Twitch](https://www.twitch.tv/devopsjacquie): the guest says the handle is "DevOps Jackie" on Twitter and Bluesky (00:42:01); Twitch is where the guest says live coding happens (00:10:15). Bluesky and Instagram handles are spoken but not in the description. - [Nebius](https://nebius.com): the guest's current employer, named at 00:01:39. The transcript does not describe what Nebius sells. - [HashiCorp](https://www.hashicorp.com): a former employer, named at 00:01:39. - [The Full Stack, season 1 (CodeTV)](https://codetv.dev/series/the-full-stack/s1): the description labels it a reality TV show; the guest describes the show at 00:09:33 without naming it. - [CodeTV](https://www.codetv.dev): the production company named at 00:09:33. - [Strudel](https://strudel.cc): the live-coding music environment named at 00:39:51. - [DevRelCon](https://developerrelations.com/devrelcon/): the conference in the host's intro (caption: "Devrocon") and discussed by the guest at 00:28:10 (caption: "Devrilcon"). - Claude Code (Anthropic) and Codex (OpenAI): tools named at 00:20:18 (caption: "cloud code") and 00:39:51. No URL is given in the transcript. - Salt (SaltStack), git, Kubernetes: tools named at 00:22:28 and 00:23:12 as early DevOps discoveries. - Jason Lengstorf and Kelsey Hightower: likely correct forms of the captions "Jason Langdorf" and "Kelsey High Tower" at 00:09:33. The description does not confirm the names; check the show page. ## 🗣️ Paraphrase The guest came to DevRel from music, DevOps and a lot of public discomfort. A fear of speaking in meetings pushed the guest onto bigger stages until the fear shrank, and a TV competition later pushed the guest into building a product from scratch. The product, Launch Buddy, turned a launch idea into hooks, evals and a messaging template, and the guest stopped it before it wrote the finished piece, because the story and the voice should stay personal. The host and guest agree that AI-written posts and pull-request descriptions have gotten subtler and still read as machine-written a few paragraphs in. The guest's answer is a split: let agents remove production toil (editing, scaffolding, dashboards), and keep the translation of engineering detail into plain English with a person. The guest also says the lower cost of building has brought back a sense of wonder that the guest felt in the first years of DevOps, and the host agrees from their own Claude Code experience. On measurement, both say the old metrics (stars, signups) no longer carry the argument. The guest watches signals further downstream, such as non-employees answering questions and whether users help other users, and runs an agent-built dashboard over posts and meeting notes. The host used Claude to build a dashboard and says much of DevRel is still vibes. The episode ends on predictions (prices rise, open weights return, audiences reward human voice) and on hobbies: Strudel, an anniversary song, keyboards and plants. ## 📰 Newsletter **A DevRel practitioner on why the human layer stays.** Episode 36 of *Community & Code* is not about WordPress, but its questions land close to home for anyone running docs, a community channel or an ambassador program. Guest Jacquie Capur (the caption spells it "Jackie Kapor") describes automating production toil and keeping the point of view personal, and says DevRel measurement now looks past stars and signups toward non-employees answering questions in a Discord. Three things worth taking from the 43 minutes: - A launch-story agent stopped before generation on purpose (00:14:29), because the story and voice are the job. - A DevRel "pyramid scheme of helping" is a scaling idea you can count: how many people your help helped others (00:31:53). - "Can robots read our docs?" (00:35:28) is a one-line audit prompt for any documentation team. Every number in the episode is self-reported, so treat it as a practitioner's view, not data. Links: [the episode page](https://communitycode.dev/episodes/episode-36-jacquie-capur/), [Strudel](https://strudel.cc) for the live-coding music tool mentioned at 00:39:51, and [CodeTV's The Full Stack](https://codetv.dev/series/the-full-stack/s1). ## ✅ Action items #### DevRel - Developer advocacy: before scripting any agent-assisted launch content, mark where the speaker's own story starts and stops, using the guest's stop-before-generation boundary at 00:14:29 as the template. - Education: run one workshop where participants turn an engineering changelog into plain English by hand first, then compare it with agent output, to teach the translation layer the guest names at 00:17:26. - Developer experience: add an "AEO check" to docs reviews: can an agent answer the top ten support questions from the docs alone (guest, 00:35:28). - Developer marketing: test the guest's observation that the same post performs differently on LinkedIn and X (00:33:19) by logging per-platform results for four weeks before changing channel mix. - Community management: trial one downstream metric from 00:31:09, non-employee answers per week in one channel, beside the current join counts. - Internal/external/cross-project: agree what attribution DevRel owes GTM before a dashboard exists; the host and guest both describe the pressure at 00:28:53 and 00:30:25 without naming a standard. #### OSPO - Policy and governance: the host says a project requires disclosure of AI assistance (00:15:12). Write down what disclosure wording contributors use, since this source gives none. - License compliance: no direct action from this source; nearest relevance: music and video tools named (Strudel, editing workflows) have their own licenses to check before reuse in public demos. - Security management: no direct action from this source; nearest relevance: a launch agent that takes a GitHub link as input (00:14:29) reads third-party repositories, which is a data-handling question worth a note. - Software inventory: no direct action from this source. - Developer support: no direct action from this source; nearest relevance: contributors may paste AI-generated pull-request descriptions, and reviewers lack a shared check for them (00:15:12). - Community engagement: the guest's prediction of an open-weights resurgence (00:38:21) is a reason to keep upstream-model dependencies swappable; note the guest's employer interest is unstated. - Strategic alignment: pick one leading and one lagging DevRel indicator and report both, since neither host nor guest gives a settled set. #### Community - Maintainers: no direct action from this source; nearest relevance: the host's confusion reading an AI-written blog post (00:16:42) is a review-quality signal maintainers can name. - Contributors: add a one-line "say it out loud" AI-assistance note to contribution guidelines, echoing the host's remark at 00:15:12. - Sponsors: no direct action from this source. - Ecosystem: no direct action from this source; nearest relevance: the pandemic-era Zoom limit versus sign-ups at 00:28:10 is a venue-planning example for cross-project events. #### WordPress - Core: no direct action from this source. - Core Performance: no direct action from this source. - Core AI: no direct action from this source; nearest relevance: the guest's use of a second model to judge a first model's output in an eval loop (00:19:35) is a small, public example of model-as-judge evaluation. - **Core Program Team:** no direct action from this source; nearest relevance: the guest's "human translator" role (00:17:26) describes what program-team communications do between engineering threads and contributors. - Mobile: no direct action from this source. - Playground: no direct action from this source; nearest relevance: a throwaway browser demo is the format of the guest's whimsy projects (00:18:54). - Community: pilot the non-employee-answers metric at 00:31:09 as a "non-affiliated answers" count in one support or Slack channel. - Training: no direct action from this source; nearest relevance: the fear-then-bigger-stage story at 00:03:49 suits a speaker-mentorship lesson. - Marketing: no direct action from this source; nearest relevance: the saturation claim at 00:25:20 is a hypothesis for editorial-calendar planning, not a measured fact. - Support: no direct action from this source; nearest relevance: "Can robots read our docs?" (00:35:28) applies to knowledge-base articles that agents will quote. - Design: no direct action from this source. - Documentation: run the AEO test on three HelpHub or DevHub pages: ask an agent each page's top question and score the answer. - Photos: no direct action from this source. - TV: no direct action from this source; nearest relevance: the CodeTV live-coding format (00:09:33) is a possible format idea for WordPress TV. - Polyglots: no direct action from this source. - Accessibility: no direct action from this source; nearest relevance: the episode's own captions garble names, which is the failure the team reviews. - 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: no direct action from this source; nearest relevance: the guest's description of speaking as exposure therapy (00:03:49) fits a first-talk mentoring track. - Projects: no direct action from this source. - Funding experiments: no direct action from this source; nearest relevance: if The WP Open Community Collective funds DevRel-style work, the downstream metrics at 00:31:09 are candidate reporting lines. ### FAIR - Technical independence: no direct action from this source; nearest relevance: the open-weights prediction at 00:38:21 argues against single-vendor model dependence. - Policies and covenants: no direct action from this source. - Productized governance artifacts: no direct action from this source. ## 🧪 Hypotheses 1. A community channel's count of non-employee answers per week predicts 90-day retention of new members better than its count of joins (guest's metric, 00:31:09). 2. Docs pages that let an agent answer their top question correctly in one pass produce fewer repeat support tickets on that topic than pages that fail the agent test (AEO claim, 00:35:28). 3. Readers detect AI-written posts within three paragraphs at a rate above chance, and a short human-written lead lifts completion on the same post (host and guest, 00:15:58 and 00:26:02). 4. A DevRel team that adds a downstream metric (non-employee answers) alongside a reach metric gets a more stable quarterly goal than a team using reach alone. 5. Identical posts published to LinkedIn and X yield different engagement per follower, with a gap above 2x for at least half of topics (guest, 00:33:19). 6. DevRel practitioners who automate editing and scaffolding report more side projects shipped per quarter than those who do not (guest and host on lowered barriers, 00:21:02 and 00:24:37). ## 📊 Methodology - **H1.** Pick one community with join and answer logs. For each cohort of new members, label answers from non-employees, then compare the correlation of 90-day retention with joins versus with non-employee answers. Report effect size and pre-register the retention definition. Source: platform exports. - **H2.** Select 30 docs pages, record their top-question tickets for 90 days, run each page through two agents with a fixed prompt, and score pass or fail. Compare ticket rates between passing and failing pages. Control for page traffic. - **H3.** Show 200 readers a mix of human-written and AI-written paragraphs from a public source; ask which paragraph, if any, reads as machine-written, and log the paragraph number. Then A/B test a human lead on one live post and measure scroll depth. - **H4.** Compare two quarters of goal-setting records from teams using reach-only and reach-plus-downstream metrics. Measure variance from target and the number of mid-quarter goal changes. - **H5.** For 20 topics, publish matched posts on both platforms, record engagement per follower at 7 days, and take the ratio. Pre-register the 2x threshold. - **H6.** Survey practitioners at two points a quarter apart on side projects shipped and hours per week spent on editing and scaffolding; use difference-in-differences against a non-automating group. Self-report is the weak link, so ask for links to shipped work. ## ❓ Questions - Which stage did the guest speak on, and what was the audience number? The intro says 1100 at DevRelCon, the guest names DevOps Toronto, and the caption lost the digits (00:04:33). - Which surname does the guest use professionally, Capur or Grindrod, and why do two description links use the second? - What exactly does "Launch Buddy" evaluate when it "tests" 20 hooks, and what counts as passing (00:13:47)? - How many non-employee answers per week is "good," and what denominator makes the number comparable across communities (00:31:09)? - Which survey does the host cite about Gen Z, and what does "most critical" measure (00:36:53)? - What is the evidence that TikTok handles more searches than Google, and for which user groups (00:32:37)? - Does the guest's employer, Nebius, sell products affected by an open-weights resurgence? The transcript does not say, and it matters for weighing the prediction at 00:38:21. - Which of the host's cited tells separate AI-written PR descriptions from human ones "a couple paragraphs deep" (00:15:58)? ## 🔮 Future research - **DevRel downstream metrics.** Little public work compares stars-and-signups to answer-based or help-chain metrics for predicting product adoption. A cross-company dataset would settle whether the guest's shift is real. - **Agent readability of docs.** AEO is named at 00:35:28 with no measurement approach; a published benchmark for docs retrievability would turn the prompt into a standard. - **Human-voice premium.** The claim that audiences will weight human-written insight more (00:26:02) is testable with controlled A/B reading studies. - **Pace of DevRel burnout under AI.** The host and guest describe pressure to output more; no data separates tool-driven relief from expectation creep. - **Open-weights adoption and pricing.** The prediction at 00:38:21 invites a tracked dataset of price changes and model availability notices. ## 💡 Implications - **Practice.** Draw a line around what stays personal (story, taste, final voice) and automate what does not (editing, scaffolding, dashboards). The guest's stop-before-generation design at 00:14:29 is a worked example. - **Policy.** Projects that require AI-assistance disclosure (host, 00:15:12) should say what a disclosure looks like; the episode treats "say it out loud" as the whole rule. - **Security.** An agent that ingests GitHub links and meeting notes (00:14:29, 00:32:37) handles third-party and internal data; access and retention need a decision before it becomes a standing workflow. - **Education.** The exposure-therapy path into public speaking (00:03:49) and the "translate geek speak" skill (00:17:26) are teachable, and both can be practiced without AI. - **Funding.** If sponsors fund DevRel, downstream metrics such as non-employee answers give a more defensible report than reach alone; the episode does not supply a proven set. - **Source literacy.** The episode is a practitioner's account with unsourced figures; treat it as a lead to test, not a citation. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch Reading Ease about 78 and Flesch-Kincaid grade about 7.2 on the 8,491-word transcript body, computed with a heuristic syllable counter and punctuation-based sentence splits. Spoken filler and caption punctuation make the sentence count rough, so treat both as plain-conversation estimates, not exact scores. #### 👀 Visual learners - Present: the format is a video interview, so the picture is two people talking; the guest refers to things in view (a dog at 00:19:35, a MIDI keyboard "over here" at 00:40:34). - Improve: chapters, on-screen name captions for the host and guest, and a one-slide metrics diagram of the guest's downstream-signal pyramid (00:31:53). #### 🔊 Aural learners - Present: the whole episode is audio-first with natural conversation, laughter and music at open and close. - Improve: add a 2-minute spoken recap and mark the three claims worth replaying (00:17:26, 00:31:09, 00:35:28). #### 🗣️ Verbal learners - Present: stories and analogies carry the content: Great British Bake Off, a castle of doors, observability for DevRel, a pyramid scheme of helping. - Improve: define GTM, AEO and "taste layer" in the description; they are used without definition (00:28:53, 00:35:28, 00:18:10). #### 🧩 Kinesthetic learners - Present: the guest describes projects the listener could copy: a dog-photo game, short-video automation, Strudel music. - Improve: link a repository or a recipe for the video workflow and the eval loop; none is given in the transcript. #### 🧠 Logical learners - Present: a clear contrast between old and new DevRel metrics (00:31:09) and a stated mechanism for scaling help (00:31:53). - Improve: the claims about saturation, pricing and TikTok search share carry no figures; add sources or label them as predictions. #### 🧑‍🤝‍🧑 Social learners - Present: the format is dialogue, the guest invites questions at 00:43:28, and the host links the guest's handles. - Improve: add a discussion prompt per segment and a place for audience replies, and list the Bluesky and Instagram handles in the description. #### 🧘‍♀️ Solitary learners - Present: the episode is long enough for deep listening and self-contained. - Improve: add reflection prompts such as "which of my metrics measure downstream help?" in the description or show notes. #### 🌿 Naturalistic learners - Present: the guest mentions about 60 plants at 00:42:01 as part of a colorful life; nothing else touches nature. - Improve: no strong fit in this source; skip it instead of forcing a tie. ## 📊 Accessibility assessment Judged from the transcript and metadata only; the video frames, player and page were not inspected. #### Screen reader semantics - Present: the description lists every link with a label (Sessionize, HashiCorp author page, Nebius, and so on). - Improve: the description text is padded with blank lines between bullets and ends with a "The post ... appeared first on" boilerplate line, which reads as noise in a screen reader (1.3.1 Info and Relationships). The description gives no one-line summary of the episode. #### Keyboard navigation - Present: the YouTube player supplies keyboard operation. - Improve: no content-level barriers are visible from the transcript; the absence of chapters (metadata lists none) forces scrubbing instead of jumping, which relates to 2.4.5 (Multiple Ways) on the hosting page. #### Contrast and readability - Present: unverifiable from a transcript. - Improve: confirm on-screen name labels, if any, meet 1.4.3 (4.5:1). #### Text resizing and reflow - Present: platform captions can be resized by the viewer. - Improve: burned-in text, if any, will not resize (1.4.4); check once. #### Media and captions - Present: the transcript exists; auto-captions mark laughter and music, which gives some non-speech context. - Improve: auto-captions garble the guest's name and most proper nouns (DevRel as "Devril," HashiCorp as "Hashi Corp," Nebius as "Nebus," DevRelCon as "Devrocon," and the names of the show's guests), which fails the accuracy intent of 1.2.2 (Captions, Prerecorded). A spoken sign-off of handles at 00:42:01 and 00:42:45 relies on listening: handles for Bluesky and Instagram are missing from the description (1.2.1 and 1.2.3 intent). Visual references ("over here," the dog in the background at 00:19:35) are not described in speech (1.2.5, Audio Description, Prerecorded). Music plays under the opening and closing speech; 1.4.2 only requires control for audio over 3 seconds that autoplays, and 1.4.7 (AAA) asks for low background audio. #### Forms and interactions - Present: none in the video. - Improve: the invitation at 00:43:28 for questions has no visible contact form; the description lists no email or link for it (3.3.2 intent, applies to the destination page). #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (inaccurate captions on every proper noun), **1.2.5** (visual references not narrated), **1.3.1** (padded description), **2.4.5** (no chapters on the hosting page). The highest-impact fix is a corrected caption file for the name list above. ## 🌍 DEIB assessment #### Cultural breadth - Present: references span a US and Canadian tech scene (Portland, Toronto), a British TV format, and early-2000s web culture. - Improve: all examples are North American or British and English-language; invite one guest or example from a non-Western DevRel community. #### Inclusive language - Present: no exclusionary terms; language is warm and casual. - Improve: "crazy" appears several times (for example the guest at 00:10:15 and 00:22:28) and the host calls themself "really dumb" at 00:16:42; both are low severity. A jab about lawmakers and "tubes" at 00:36:53 is political shorthand and could be dropped. Generational labels ("millennials," "Gen Z") are used as groups with shared habits without data (00:32:37). #### Neurodiversity support - Present: both speakers normalize losing a thought (00:26:02 and 00:31:09) without shame, which models a low-pressure conversation. - Improve: no chapters, long free-flowing segments and overlapping [laughter] and [music] cues make navigation harder; chapters and a segment list would help. #### Representation - Present: two voices, host and guest, with the guest describing a musical background, a spouse in SRE, and a pathway from side hobbies to a career. - Improve: no third voice, and all voices are of working practitioners at large companies; adding a community-run-project perspective would widen the lens. #### Psychological safety - Present: when the host loses a thought, the guest answers "okay, it happens" (00:26:02), and the guest admits a blank of their own at 00:31:09; the guest admits fear and nerves publicly. - Improve: no content warnings are needed; a standing note on how the show handles guest corrections would help guests who spot a caption error. ## ✅ Logical validity #### Strong reasoning - **Mechanism, 00:17:26.** The claim that a human still has to translate engineering detail is backed by the guest's own role: translating "geek speak" from engineers or from AI. The host adds that trust comes from knowing the writer (the "taste layer"). The mechanism is specific. To strengthen it, give one before-and-after example. - **Design constraint, 00:14:29.** The guest stopped the agent before generation and kept the checklist. That is a concrete rule with a reason (voice and story stay personal), not a slogan. - **Comparison within experience, 00:24:37.** The guest separates two periods of "do more with less" pressure (2020 to 2021 versus now) and names what changed (tools that take chunks out of the work). It is a before-and-after within one person's career, which suits a first-person claim. - **Scaling arithmetic, 00:31:53.** If each person helped helps two more, reach doubles per hop; the guest uses it to argue one-to-one DevRel does not scale. The logic holds if the second hop actually occurs, which is the thing to measure. - **Self-checking on pace, 00:23:56 to 00:24:37.** The host pushes back on whether anyone is less rushed, and the guest answers with a distinction between tool relief and expectation pressure instead of agreeing. #### Fallacies - **Unsourced statistic, 00:32:37.** "Tik Tok was getting more searches than Google." No source, date or population. Fix: name the study and the user group it covers. - **Appeal to an unnamed survey, 00:36:53.** "a bunch of Gen Z's were were uh surveyed." The host does not name the survey, sample or question. Fix: cite it, or label it as hearsay. - **Hasty generalization, 00:36:11.** The host's switch back to plain search is offered as part of a trend ("I feel like this is this is a thing that that we're seeing kind of trend as well"). One person's habit is not a trend. Fix: say "I switched; I don't have data on others." - **Anecdote as trend, 00:25:20.** "what I'm starting to see now is people are kind of pulling back" rests on what the guest sees online. Fix: say it is an observation and name the platforms. - **False analogy risk, 00:38:21.** Rideshare and delivery pricing is used to predict AI pricing ("already seeing the price starting to go up"); the markets differ in cost structure and competition, and no price data is given. Fix: cite price changes and keep the analogy as illustration. - **Inconsistent figure, 00:00:13 versus 00:04:33.** The intro says 1100 strangers at DevRelCon; the guest names DevOps Toronto and the caption drops the number. This is not a reasoning error by the guest, but it is a conflict inside the source. Fix: confirm which stage the figure belongs to. - **Unfalsifiable hedge, 00:17:26.** "I don't feel like we are at risk of losing our jobs to to AI yet." The word "yet" and "feel" mean no outcome can contradict it. Fix: state what evidence would change the view. #### Tone of voice - Warm, casual and self-deprecating: laughter appears throughout, the host jokes about their own forgetfulness at 00:26:02, the guest jokes at 00:01:39 about having been in DevRel "for a long time" and then says it was a joke. The register is conversational and sincere; the guest's strongest language is saved for wonder and for what the guest wants to keep human. - Opinions are flagged as opinions ("I feel," "I think"), which fits an interview and is also why most claims here are predictions, not evidence. The host's sentences are longer and more hedged than the guest's, and both interrupt each other with agreement, which reads as rapport. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a response to an accusation, and this is a friendly interview with no accusation or defense. For completeness across the three prongs: **Deny**: no denial occurs; the guest admits blanks and nerves (00:03:49, 00:31:09). **Attack**: no one is attacked; the host's jab at lawmakers at 00:36:53 is a general political aside, not aimed at an accuser. **Reverse victim/offender**: no such move appears; the nearest thing is the guest's joke that the "pyramid scheme of helping" quote will be taken out of context (00:31:53), which is humor, not reversal. ## 🎯 Conclusion The durable ideas are small and specific. First, the guest's boundary for using agents: automate the toil and the scaffolding, and keep the story and voice with a person (00:14:29, 00:17:26). Second, a measurement direction for DevRel: look past joins to whether non-employees answer questions and whether helped users help others (00:31:09, 00:31:53). Third, a check for any docs team: can robots read your docs (00:35:28)? The weaker parts are the unsourced figures and trend claims: the TikTok-versus-Google search claim, the Gen Z survey, rising AI prices, and a content-saturation backlash all come from the host's and guest's experience, not data. The intro and the guest's own account also disagree on the speaking milestone, and the guest's surname appears two ways across the description. **Next steps for readers.** Confirm the guest's surname and the garbled names against the episode page before quoting anyone. Try the non-employee-answers metric in one community channel for 30 days, and run the AEO test on three docs pages. Treat the predictions about prices and open weights as hypotheses to track, and note that the guest's employer interest is not disclosed in the transcript. ## 🔄 Refactor Appendix — Episode 36: Jacquie Capur > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied: travel anecdotes, hobbies and the reality-show logistics were dropped because they carry no claim a note would reuse. > > 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] DevRel translates engineering detail into plain language > [!quote] [00:17:26] > taking the the geek speak from the engineering team and translating it into human English **Context & use**: - 00:17:26, the guest's answer to why DevRel jobs are not at risk from AI yet; the host agrees and adds a "taste layer." - The claim is the guest's self-description of a core DevRel skill. It is a practitioner's view, not evidence about job risk. **Tags**: #xref #aphorisms <!-- atomicity: 1 idea ✓ --> #### [H2] Automating toil frees room for better ideas > [!quote] [00:18:10] > my happy path is like it's very similar to DevOps like automate away the toil. **Context & use**: - 00:18:10, setting up the guest's whimsy examples (terminal visuals, dog game, video workflow). - This is the guest's design rule for AI use. Pair it with the stop-before-generation boundary at 00:14:29 when recording it. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H3] Early DevOps felt like opening doors in a castle > [!quote] [00:22:28] > I open a door and I'm like oh configuration management. **Context & use**: - 00:22:28, the guest's example of "wonder" after the host names it; configuration management with Salt after hand-updating about 142 VMs. - The metaphor explains the pattern the host also reports: a lower barrier from idea to prototype (00:21:02). **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H4] DevRel health shows in non-employee answers > [!quote] [00:31:09] > it's not just how many people joined our community, it's how many non-emp employees answered questions in the Discord, right? **Context & use**: - 00:31:09, the guest's answer to the host's question about valuable DevRel metrics; caption "non-emp employees" is probably "non-employees." - This is the most reusable metric idea in the episode and the guest gives no threshold or data, so it is a hypothesis to test. **Tags**: #xref #share #articleseed <!-- atomicity: 1 idea ✓ --> #### [H5] Open-weights models will gain share > [!quote] [00:38:21] > I do think we're going to see a resurgence of open source. **Context & use**: - 00:38:21, the guest's prediction after noting rising prices and short-notice changes to proprietary models. - Prediction only. The guest's employer is named at 00:01:39 and the transcript does not say what it sells, so weigh the claim accordingly. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> ### 🧠 Atomic Split Pad #### [P1] Automate the toil, keep the voice **Claim** (one sentence, Courtney's voice): Agents belong in the production steps of communication work, and the point of view stays with the person whose name is on it. **Evidence pointer**: [[#H1]] · [[#H2]] **Scope & conditions**: Holds for personal-voice work such as DevRel posts, talks and docs lead paragraphs. It weakens for reference material where no voice is expected, and a team still needs a rule for disclosing assistance. **Links**: - [[Human-in-the-loop AI writing]] — because extends - [[AI-assisted contribution disclosure]] — because applies-in open-source projects - [[Automation as pipeline]] — because contrasts on where the human stays <!-- atomicity: 1 idea ✓ --> #### [P2] Count downstream help, not joins **Claim** (one sentence, Courtney's voice): A community's health shows in how many people outside the company answer other people's questions. **Evidence pointer**: [[#H4]] **Scope & conditions**: Holds where the channel has public answers and a way to tell employees from non-employees. It fails in a new community with few members, and it is a hypothesis until tested against retention. **Links**: - [[DevRel metrics]] — because extends - [[Community health indicators]] — because applies-in Discord and Slack communities - [[Stars and signups]] — because contrasts on reach versus help <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Confirm the guest's surname (Capur versus Grindrod) and the garbled names (Lengstorf, Hightower, the AWS cast member) against the episode page and show page: https://communitycode.dev/episodes/episode-36-jacquie-capur/ #fleeting #inbox - [F2] Resolve the speaking-milestone conflict: the intro says 1100 at DevRelCon, the guest names DevOps Toronto and the caption lost the number (00:04:33) #fleeting #inbox - [F3] Find the source for "Tik Tok was getting more searches than Google" (00:32:37) before repeating it #fleeting #inbox - [F4] Open question: what does a useful AEO test for a docs page look like beyond "can robots read our docs" (00:35:28)? #fleeting #inbox - [F5] Article seed: stop an agent before generation on purpose, as the guest did with Launch Buddy (00:14:29) #fleeting #inbox #articleseed ### → Things 3 (max 2 actions) - [ ] Verify the guest's surname and the garbled names against communitycode.dev and codetv.dev before citing this episode: 10 min #things - [ ] Trial a non-employee-answers-per-week count in one community channel for 30 days: DevRel metrics #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Episode 36 Jacquie Capur.md|Episode 36: Jacquie Capur by Jacquie Capur (guest; the transcript captions spell the name "Jackie Kapor"), Chris Reynolds (host)]] ## Highlights - [00:17:26] I don't feel like we are at risk of losing our jobs to to AI yet. Um because there I feel like there does still need to be like a human translator like you know >> I I'm good at this job because I have always been really good at like taking the the geek speak from the engineering team and translating it into human English and then communicating with with people and whether I'm taking that from engineers or I'm taking that from AI like it still needs that human translation layer right? >> Yeah. And then there's that like you know that experience that taste layer where it's like I trust that you know maybe I trust that Jackie wrote it because Jackie's been in platform engineering for so long and also is kind [00:18:10] of vocal about making sure that we only see things that I believe in. Right. Yeah. >> So I think that layer um goes missing when you're just only relying on like an agent to do it. >> I I think my happy path is like it's very similar to DevOps like automate away the toil. I think when we take the toil out of the things that we do, but we enable it to like let us, you know, pursue our whimsy, pursue our like better ideas. I think that's where I'm happiest with using AI. >> Yeah. What whimsical ways are you using AI today? [laughter] >> Yeah, fair. Um I think there's a lot of demos in the last like we'll just say this year. There's been a lot of demos this year that I don't think I would have been able to pursue in the past without it because maybe that means ([View Highlight](https://read.readwise.io/read/01m564fk92yc5ky100k20j5cpm)) - Note: [H1] DevRel translates engineering detail into plain language - Tags: atom/literature/h1 - [00:21:46] >> Oh yeah. Oh, I forgot about those. [laughter] Um, so that was that was how I learned. And like back in those days, like you just like, "Oh, I see like this cool thing on this website. I'm just going to look at the source and copy that and put it in my thing and tweak a couple things and then magically it's it's it's there." And it's like >> there's a little bit of that sort of like uh discovery and and >> dare I say wonder uh when you're when you're able to sort of make ideas happen um like very very quickly. >> Yeah. No, I think Wonder is the correct word for it. Like when I first took that original um DevOps internship job, basically it was a lot of like it kind of felt like you're like walking through [00:22:28] this castle and like every new door had a new exciting thing behind it. Like you know I open a door and I'm like oh configuration management. Wow, look what we can do with you know um salt. Salt is what I was using at the time. So I was like look what we can do with salt. Like this is crazy. Did you know you could just do these things? It just like I had just updated I want to say it was like 142 um like VMs like manually you had to go in things and >> took a while and then I discovered config management and then it just kind of kept being like that like you know then I learned about git and that was great that was awesome [laughter] you know there's a whole bunch of things in there where learned about pipelines and I was like whoa that's great I don't have to just script everything and hope it will >> work everywhere even if like the file changes or the user changes or ([View Highlight](https://read.readwise.io/read/01m564fnhhem64s5s62bg6t1ct)) - Note: [H3] Early DevOps felt like opening doors in a castle - Tags: atom/literature/h3 - [00:30:25] much of Devril is vibes like >> you know like it's it's it's >> putting out into the world something positive to to earn goodwill and like support of the developer community and then presumably those developers will want to work with me right like or want to work with the you know um and how do you like put numbers on that and so I've been like I it's a problem that I gave to Claude uh and to build out a a dashboard of like stuff that we can measure because I felt like we needed to represent this. Um, but I'm curious because we're talking about GTM, we're talking about uh how to quantify the work that we do. Like what are what are things that like [00:31:09] you consider valuable sort of metrics to score devril efficiency or e effect efficacy? Yeah, I feel like that's like man that's tough like [laughter] because I think like if you'd asked me three years ago I could have been like GitHub stars and signups and you know the answer was so easy then and it's just not that anymore. Like now we're looking at that like second or third layer of things which are things like >> you know um it's not just how many people joined our community, it's how many non-emp employees answered questions in the Discord, right? it's >> like I'm I'm drawing a huge blank now. I'm like I have nice metrics that I like tracking, but they're uh not present. ([View Highlight](https://read.readwise.io/read/01m564fqx1gb65gwa6mqryajk0)) - Note: [H4] DevRel health shows in non-employee answers - Tags: atom/literature/h3 - [00:37:38] a dump truck. uh sort of like politicians in in uh Washington DC that have no idea what they're talking about and they're making policies for everyone else based on these like really flawed uh assumptions. [snorts] >> Yeah, [laughter] I don't know. I think I agree like it's hard to give feedback on something when you're not using it. And so I definitely think that's a good aspect of people using it. But um I don't have a more complex thought on that right now. [laughter] Like I know it's it's going to be like, you know, the original Ubers and Door Dashes where the price is going to be really easy and accessible and then you get used you kind of get accommodated to that lifestyle or those things that you now have available to you and we're [00:38:21] already seeing the price starting to go up. So I think it's going to be >> interesting what comes next. Um I don't have too much on like the policy side of things, but I do think we're going to see a resurgence of open source. like we're going to start seeing more open weights, open models. I think that's going to become more popular as we're seeing the, you know, limited notice changes that are being made to proprietary models. And you can't, you just kind of can't have a tool that may not be as available tomorrow as it is today because suddenly prices have changed or >> right, >> you know, these like 24-hour price cycles when it used to be months, quarters, years, >> right? >> It's interesting. >> Yeah. Yeah. >> Yeah. ([View Highlight](https://read.readwise.io/read/01m564ft89j0wy0nqenccx6z17)) - Note: [H5] Open-weights models will gain share - Tags: atom/literature/h3