Talk by Priscila Andre de Oliveira (Sentry) on using AI primarily for code comprehension rather than code generation. Drawing on an analysis of 116 of her own Claude sessions from daily work (67% comprehension, 2% code generation), she explains how a 15-year-old codebase serving ~100,000 organizations rewards understanding over generation. She describes her personal "catch me up" skill with six exploration modes — architecture, conventions, feature traces, syntax, testing, and history — and the workflow of verifying what the agent found before letting it plan and implement, arguing that a misaligned mental model is the source of low-quality output. Timestamps cover Sentry's engineering scale, AI-driven projects, code quality and technical debt, AI usage patterns, the skill's architecture, a demo, and planning vs. implementation workflows. # Video Identity Observation — Channel Presentation Evidence ## Video Overview A conference-style **sponsored talk** by Priscila Andre de Oliveira, a senior software engineer at Sentry, about how she uses AI (particularly a custom "catch me up" comprehension skill) in her daily work on Sentry's large codebase. She closes with an explicit sponsorship pitch: > "Try Sentry. This was a sponsored talk. We have a booth downstairs... If you scan this QR code, you're going to get a three months free trial of our business plan." ## Intended Audience - **Developers working in large, complex codebases**, especially those experimenting with AI-assisted workflows. Evidence: discussion of AI agents, Claude Code, prompting, PR review, and the framing "That's why you all are here" — indicating a live developer-conference audience. - **Practitioners of JavaScript/TypeScript tooling** — the speaker identifies as co-organizer of **Vienna JS** ("a very traditional meetup in Vienna and we talk all about JavaScript"), suggesting the talk was delivered at a JS-oriented event. - Audience is presumed familiar with AI tooling concepts ("I think you guys also heard about it outside like even cloud code has this planning mode"). ## Claimed Scope (Self-Description) - The speaker, not the channel, provides the scope framing: a **personal workflow talk** — "today I'm going to share with you how I use AI at Sentry and this [skill] I use the most in my day-to-day work." - Includes **employer scope**: Sentry presented as "a full observability platform" with error monitoring, metrics, profiling, and "agentic tools" — per the speaker's description, not independently demonstrated. - Includes **community credentials**: maintainer of **Verdaccio** ("an open-source NPM registry"), co-organizer of **Vienna JS**, six years at Sentry. ## Observable Presentation Patterns - **Educational/demo format**: the talk includes a live demo of a custom AI skill ("Catch me up on how this repository works...") and quantified self-analysis (Claude analyzed "116 sessions" and classified usage into six categories; "67% of my AI usage was comprehension and only 2% code generation"). - **Promotional closing**: the talk explicitly functions as marketing for Sentry's product (free-trial QR code, booth reference), suggesting the channel/event hosts sponsored corporate talks. - **Opinionated practitioner perspective**: recurring themes of comprehension over generation — "the biggest unlock from AI in a large code base isn't generation. It's comprehension" — and quality advocacy: "don't ship slop code into the code base that pays your salary." - **Humor and self-disclosure style**: jokes about being an "agent manager," "this is my vibe coded presentation," and personal context (Brazilian in Vienna, mother of a toddler) — a personable presentation style. ## Creator Claims vs. Demonstrated Evidence - **Claim (unverified)**: "since December 2025, I haven't coded anymore. I'm only prompting." Stated by the speaker; not independently verifiable from the material. - **Claim (unverified)**: Sentry statistics — "founded in 2010," "400 employees," "100k organizations depending on this code base," "~100 PRs merged every day." All are speaker statements. - **Demonstrated**: her methodology (creating a local skill defined as "just... an MD file with human language," running it live on an unfamiliar repository) is shown in the demo portion. - **Demonstrated**: the talk's promotional intent, evidenced by the closing sponsorship callout. ## No Curated Metadata Curated metadata was not supplied for this video; identity observations rest solely on the title and transcript. # Delivery Observations — Priscila Andre de Oliveira, "The AI Skill I Rely On Daily" ## Tone & Overall Manner - **Conversational and humorous throughout.** The speaker jokes early and often: after noting her official title, she says, *"but I have given me a little promotion and I am now an agent manager. Um no salary raise, but at least my reports they don't complain."* - **Self-deprecating and candid**, especially about her own surprise at her data: *"And guess what? This impressed even me."* and *"So, this I was very surprised."* - **Audience-interactive warmth**, with direct addresses: *"Can you guys see this well?"*, *"I don't know if you guys have the same"*, and a joke about her slides: *"By the way, this is my presentation running."* ## Verbal Habits & Fillers - **Heavy use of "uh" and "um"** as connective fillers across the entire talk, e.g., *"Um so but before we dive into that, let me tell you who I am."* - **Frequent "Yes" / "Yeah" as transitions and emphasis markers**: *"Yes, um and even this presentation was created by AI skill."*, *"Yeah, I'm not going to read it, but this information..."*, *"So, yes."* - **"So" as a recurring discourse pivot** to restart or resume thoughts: *"So, and this is Sentry."*, *"So, back to my vibe coded presentation."*, *"So, there is no dumb questions."* - **"You know" as an involvement marker**: *"maybe AI understood... understood wrongly, you know? And you need to also understand... You know?"* ## Speech Disfluencies (Noticeable Pattern) - **Self-corrections mid-word or mid-phrase** occur repeatedly: - *"the indus- uh sorry, uh this is how I feel"* - *"I'm core orga- co-organizer of Vienna JS"* - *"a few wee- a few weeks ago"* - *"I would have to open a PR uh oh open GitHub, sorry"* - This is a **stable pattern**, not an isolated moment, though it does not impede comprehension. ## Vocabulary & Recurring Phrases - **"Slop code" as a motif**, appearing twice with a rhyming closer: *"I don't want to just uh ship slop code"* and the closing *"don't ship slop code into the code base that pays your salary. Ship keynote code."* - **"Comprehension" as the keyword concept**, repeated for emphasis near the end: *"the biggest unlock from AI in a large code base isn't generation. It's comprehension."* - **Repetition of key statistics for reinforcement**: *"67% of my AI usage was comprehension and only 2% code generation"* — restated later as *"67% of my prompts are com- comprehension and only 2% generation."* - **AI-assistant anthropomorphization**: refers to Claude as *"my favorite teammate Claude"* and *"AI is the teammate who never gets tired of your questions... It's the cheapest senior engineer out there."* ## Teaching / Presenting Habits - **Personal introduction before substance**: spends the opening on biography (Brazil, Vienna, toddler, Verdaccio, Vienna JS, Sentry) before the main topic. - **Concrete enumeration** when describing tools and concepts — lists hackathon projects by name (Abacus, Warden, Junior) with one-line explanations each, and lists six exploration modes: *"Architecture, convention, feature trace, syntax, testing, and history."* - **Definition-by-simplification** for technical concepts: *"a skill is just a very detailed prompt, right? With a very clear goals."* - **Live demo embedded mid-talk**, with preparation hedging: *"I already run this skill because I don't know maybe I would have some issues."* - **Quote-and-endorse structure** when citing others: quotes Armin Ronacher, then adds *"And yes, I agree."* - **Explicit takeaway summary** at the close, framed as advice: *"align your mental model before you prompt."* - **Direct sponsorship disclosure** at the end: *"This was a sponsored talk."* ## Uncertainty Notes - The audience reactions (**laughter**, **applause**) are marked in the transcript, confirming the humor lands, but tone of voice itself (pace, pitch, volume) is **not observable** from transcript alone. - Filler frequency and self-corrections are clearly stable; whether the demo segment ran smoothly beyond the shown text cannot be verified from the transcript. # Observation Document: "The AI Skill I Rely On Daily — Priscila Andre de Oliveira, Sentry" ## Speaker **Priscila Andre de Oliveira** — self-identified as a Brazilian based in Vienna, Austria; maintainer of Verdaccio; co-organizer of Vienna JS; senior software engineer at Sentry (6+ years). Single-speaker presentation. ## Personal Context & Framing - Jokingly calls herself an **"agent manager"** — "no salary raise, but at least my reports they don't complain." - Claims that **since December 2025 she has not coded, only prompted**, including this presentation ("I didn't do anything"). - Reports creating PRs "together with my favorite teammate Claude" — bug fixes, features, refactors, cross-repo contributions. "It's real. It's working." ## Sentry Context (as described) - Full observability platform: error monitoring, metrics, profiling, agentic monitoring tools. - Founded 2010; ~400 employees; ~100k organizations depend on the codebase; ~100 PRs merged daily; fair-source with external contributors. - Position: "I don't want to just ship slop code... It's a serious business." - Internal AI tools from a hackathon: - **Abacus** — tracks internal AI usage. - **Warden** — code review agent for PRs. - **Junior** — Slack bot that analyzes threads and creates PRs to fix reported bugs. - **AI SDK testing repository** — pre-hackathon; where her team told her to *only prompt, not code*. - **Quality quarter**: three months dedicated to reducing technical debt (removing TypeScript `any`s, TODOs, unused feature flags, simplifying code). ## Core Claim: Comprehension over Generation - She had Claude analyze her cache: **116 sessions** classified into six categories (comprehension, modification, process, review, generation, other). - Result: **67% comprehension, only 2% code generation** — "this impressed even me." - Key takeaway stated: *"The biggest unlock from AI in a large code base isn't generation. It's comprehension."* ## The "Catch Me Up" Skill - A local, personal skill ("just a very detailed prompt... an MD file with human language") structuring comprehension questions into six exploration modes: **architecture, convention, feature trace, syntax, testing, history**. - Visual output (organograms, tables) because she is "a very visual person." - Demo: used on the unfamiliar AI SDK testing repo ("I am a new contributor. Catch me up on how this repository works..."). - Also used to **build context for reviewing colleagues' PRs**. - She distinguishes: vibe-coded side projects are fine, but "this is my work... I don't like to ship something I don't understand." ## How AI Made Her Faster - Incident tracking: previously GitHub/git blame archaeology; now a prompt returns answers in seconds. - Product decisions: previously waited on Slack answers across time zones; now asks AI directly. ## Cited External Positions - **Jack Nation** (blog post "Vibe Coding Our Way to Disaster," drawing on Rich Hickey's "Simple Made Easy"): proposes three phases — **research, planning, implementation**. Priscila *agrees but adds a missing step*: you must **understand the research your agent did** and steer it, possibly correcting its misreadings of the codebase. - **Armin Ronacher** (Flask creator, former Sentry): quoted — *"When more and more people tell me they no longer know what code is in their own code base, I feel like something is very wrong here."* Priscila: "I agree." ## Advice to Audience - Track your own AI usage; you may be surprised. - "AI is the teammate who never gets tired of your questions... the cheapest senior engineer out there." - "Align your mental model before you prompt" — then code flows naturally. - "Don't ship slop code into the code base that pays your salary. Ship keynote code." ## Qualifications & Tensions Preserved - She repeatedly emphasizes speed gains **are subordinate to understanding** — AI may "understand wrongly" and needs steering; her productivity claim is coupled to this caveat. - Humor/tension between "vibe coded presentation" and insistence on not shipping un-understood code at work. - Self-described surprise at her own metrics — she presents the 67/2 split as a finding, not a presupposition. - Talk ends with an explicit sponsored-talk disclosure (3-month Sentry Business trial). # Video Attribution & Provenance Record **Video ID:** `01a15ca9-3d9c-7c81-92f9-3bcbd0a676f1` **Title:** The AI Skill I Rely On Daily — Priscila Andre de Oliveira, Sentry ## Speakers & Roles - **Priscila Andre de Oliveira** — **Primary speaker / presenter** - Self-identified: Brazilian based in Vienna, Austria; senior software engineer at **Sentry** (~6 years tenure) - Additional self-identified roles: maintainer of **Verdaccio** (open-source NPM registry), co-organizer of **Vienna JS** meetup - Notes she gave herself the informal title "agent manager" (her own framing, not an official role) - **Dominik Dorfmeister** — **Colleague (mentioned only)**; credited with taking a photo of the speaker - **Daniel** — **Sentry colleague (mentioned only)**; referenced as enjoying the "Junior" bot - **Audience** — reactions only ([laughter], [applause]); no speaking lines ## Referenced Sources & Third-Party Claims - **Jack Nation** — blog post *"vibe coding our way to disaster"*, drawing on **Rich Hickey's** "Simple Made Easy" philosophy; proposes research/planning/implementation phases (speaker cites, partly agrees, adds a step) - **Armin Ronacher** — creator of Flask, described as former Sentry employee, current startup founder; quoted: *"When more and more people tell me they no longer know what code is in their own code base, I feel like something is very wrong here."* - **Sentry internal tools (described, not independently verified):** Abacus (AI usage tracking), Warden (code review agent), Junior (Slack bot creating PRs), AI SDK testing repository - **Speaker's own claim:** Claude analyzed her cache of 116 sessions; she reports 67% comprehension, 2% code generation - **Sentry statistics (speaker's own figures, unverified):** founded 2010, ~400 employees, ~100k organizations, ~100 PRs merged daily, "fair source" model ## Promotional Segments ⚠️ - **Closing segment explicitly flagged as sponsored:** speaker states *"This was a sponsored talk"* and promotes a QR code for a **three-month free trial of Sentry's business plan**; directs audience to a booth downstairs. - Sentry product overview (error/performance monitoring, metrics, profiling, agentic tools) functions as promotional context. ## Availability & Limits - No curated metadata supplied; no comment availability information. - Timestamps absent from transcript. - No independent verification of cited statistics, tool descriptions, or blog post claims possible from this source. \>\> Hello everyone. Uh today I'm going to share with you how I use AI at Sentry and this queue I use the most in my day-to-day work. Um so but before we dive into that, let me tell you who I am. Uh my name is Priscilla. I'm a Brazilian-based in Vienna, Austria. I'm a mom of a 2-years-old very energetic toddler. Uh I'm a maintainer of Verdaccio, uh an open-source NPM registry. I'm core orga- co-organizer of Vienna JS, a very traditional meetup in Vienna and we talk all about JavaScript. And uh I'm I'm a senior software engineer at Sentry. Yesterday someone told me that I don't look like an a a uh software engineer, but guess what? I am. \>\> \[laughter\] \>\> Um So my my title my official title is senior software engineer, but I have given me a little promotion and I am now an agent manager. Um no salary raise, but at least my reports they don't complain. Yes, uh this was me at work uh a few wee- a few weeks ago. Uh my colleague uh Dominik Dorfmeister found it funny to see me managing a couple of agents uh and took this picture. \>\> \[clears throat\] \>\> Um Yeah, luckily I have uh three monitors, so that works pretty well. Um this is my new reality. So the indus- uh sorry, uh this is how I feel actually orchestrating a bunch of agents. Uh yeah, it's it's weird, but it's fun. And the industry is changing, right? That's why you all are here and I'm also adapting. Uh since last year since December uh 2025, I haven't coded anymore. I'm only prompting. Yes, um and even this presentation was created by AI skill. I have Yeah, I didn't do anything. Uh \>\> \[laughter\] \>\> Um So as you can see, these are some of my recent contributions to Sentry uh and I created a a few PRs together with my favorite teammate Claude. And it's not just uh bug fixes, you know, it's also features, refactors, cross repositories contributions. So it's real. It's working. Um So and this is Sentry. Um maybe you don't know Sentry, but we are very well known for error and performance monitoring, but we we have grown into a uh full observability platform. So we have error monitoring, we have a metrics, we have profiling, we have also uh agentic tools like uh to monitor your agentic platforms. Um Yes, uh the code base is very complex and uh it was founded in 2010. It has 15 plus years of code. We got around 400 employees around the globe. We have 100k organizations depending on this code base working every day. And as employee I also depend on this code base working because I get my salary from it, right? So I don't want to just uh ship slop code. Um yeah, so it's a serious business. And we have code as well at Sentry. Uh recently we had a hackathon uh where we could we had a few days to just uh get ourselves familiar with AI and try out new things and a lot of good projects came out of this uh week, this hackathon project. And uh we have a for example Abacus. This was created to track uh the usage of AI internally at Sentry. Uh we have Warden. This is a a code review agent. Uh you can have it in your PRs. Um we have a Junior. Junior is a bot we have in our Slack uh and because usually people they see like uh Oh, uh I don't like this UI. Something changed. Uh can you go fix it or why this was changed? They like to to go and share something maybe some bug they found in Slack and then you can we can just a trigger Junior and Junior can analyze that thread and create a PR and I'll go and fix the bug. And people are having a lot of fun with Junior. Daniel can Yeah, he loves that. And uh there is also this AI SDK testing repository. Uh this was created before this hackathon, uh but this is basically a repository where we create tests to for our AI integrations. So and this was really weird because I started contributing to this repository and uh uh my team told me I shouldn't code at all. I should only prompt until I would get a nice result. So it was like a a different experience. Uh but yeah, it's working. We are using all of these tools like internally every day. And yeah, at Sentry we are going all in AI. But quality still matters. Um So uh also at Sentry last year uh during three months uh we we used this time like a quality quarter. We used this time just to improve our code base. So remove all the any types from TypeScript for example or all the to-do because usually I don't know if you guys have the same, but we had a lot of to-do uh do something else at certain time and uh so we used this time to simplify code and uh have uh remove unused feature flags and really have our our code base in a in a good shape and this is very important, right? We want to that. This is uh called it technical debt. Um So and as I told you before, the code base the Sentry code base is very complex and it's a moving target. We have about 100 PRs merged every day. We have a four offices. Uh Sentry is fair sources. So we you can contribute to Sentry. We have also contributors. Um and we are all the time deprecating components, adding new components, adding new lint rules. Uh I don't know, you name it. Like all the time something happens at Sentry. I'm there over six years now and I can go on vacation. I come back and maybe my PR uh is full of conflicts and I have to solve those conflicts and I really need to understand like all the time I need to align and and understand something. It's a daily practice. So and this is not new, right? Like uh you guys know there are the studies behind it like 70% of a uh you you may think like you just tell AI to go explore the code base and after do something, but like maybe AI understood uh not the the correct thing, under- understood wrongly, you know? And you need to also understand because maybe you need to steer the AI to go on the correct path. You know? And yeah. So This is how I'm using AI. Uh it made me faster, but not the way you think maybe because like before um let's say um an incident happened and I would have to track that down. I would have to open a PR uh oh open GitHub, sorry. I go git blame and then trying to understand like where the regression happened and now I can just prompt a simple phrase and I have it in a few seconds. Or before maybe a product product decision. Um why this changed? And then I would I don't know uh ask this question in Slack. Maybe my colleague is in another another country, another time zone and I would need to wait for that answer in the next day and now I can just uh ask AI and I have it. So it's been really useful and this made me really productive uh but but like it the understanding part of it like Yeah. And my prompts they kept repeating. Uh so I had this idea to let to to let Claude analyze my cache and see uh so it analyzed it 116 sessions and it classified everything in six categories. Comprehension, modification, process, review generation, and other. And guess what? This impressed even me. Uh oops. So, 67% of my AI usage was comprehension and only 2% code generation. So, this I was very surprised. \>\> \[laughter\] \>\> So, because my prompts kept repeating, repeating, I created a skill for me. This is skill it's locally in my computer. I could share it with someone if I wish, but I use this is for me here and it's called catch me up. Um that is structures those comprehension questions in two six exploration modes. Architecture, convention, feature trace, syntax, testing, and history. So, a skill is just a very detailed prompt, right? With a very clear goals. I can actually do like this here. Uh you can see how how is it. But it's it's just like it's a an MD file with human language. And I am a very visual person. I work at Sentry a lot on the front end part of it. And I like to see things to understand. So, uh this skill brings me like uh the organogram like the structure a table for me to understand. I think it helps a lot. And I can now give you a short demo. Um Just a minute. \>\> \[clears throat\] \>\> Um By the way, this is my presentation running. \>\> \[laughter\] \>\> Oh, it's here actually here. Um I already run this skill because I don't know maybe I would have some issues, but uh Do you remember that project I told you that I should only prompt and don't code anymore? So, in the beginning I was not familiar with that project. It was a new repository for me. And I I used my skill for that. I said, I am a new Can you guys see this well? Yeah. So, I said, I am a new contributor. Catch me up on how this repository works and clarify what it simulates a Sentry envelope and intercepts it during tests. Um I am using cloud as you can see opus uh and it gave me it gave me here a summary. And this I like this flow like how it works. Um And here also answered my question like does it simulate envelopes? No, intercepts real ones. They spawn collector. Yeah, I'm not going to read it, but this information all of this it's very useful. Like if I didn't have AI, I would have to do this myself, you know. Um I mean I don't like to to just ship something I don't understand. If it's a vibe coded project, that's fine, but this is a real serious business. It's my work. Um So, yes. This skill is helping me a lot with that and also to review PRs because maybe I'm reviewing a PR of a colleague. I have a lot of context, but not enough to approve that PR. And I wanted to have that context. So, I use this skill to give me that. Uh okay. So, back to my vibe coded presentation. \>\> \[laughter\] \>\> So, \>\> \[clears throat\] \>\> Uh Jack Nation's wrote a blog post called it vibe coding our way to disaster drawing on Rich Hickey's simple made easy philosophy. He proposes three phases: research, planning, and implementation. I think you guys also heard about it outside like even cloud code has this planning mode, right? Um so, yes. I agree with all of that, but I think it's missing this step like you need to understand the research your agent did, you know? You need to understand that and to steer as I said maybe it's going the wrong direction or maybe you need to explore something else and you need to have that to understand that. Then after that you can say, "Okay, plan that for me and let's do the implementation. Let's go ahead." Um So, Armin Ronacher he's the creator of Flask and a former Sentry. Now he's working on his own startup. Today he's going to give a talk here by the way. He's around. Um So, he wrote in his blog post, "When more and more people tell me they no longer know what code is in their own code base, I feel like something is very wrong here." And yes, I agree. Um so, what I hope you you can take away from this presentation is that the biggest unlock from AI in a large code base isn't generation. It's comprehension. I tracked my own usage and I was surprised. 67% of my prompts are com- comprehension and only 2% generation. Maybe you track your own AI usage as well and you can improve it, right? Uh so, AI is the teammate who never gets tired of your questions. So, there is no dumb questions. It's the cheapest senior engineer out there. So, just go for it. Yes, and align your mental model before you prompt because you know the code is going to flow naturally and don't ship slop code into the code base that pays your salary. Ship keynote code. Really. This is the term the industry is using. And yes, thank you. Um Try Sentry. This was a sponsored talk. We have a booth downstairs if you wanted to stop by to say hello. If you scan this QR code, you're going to get a three months free trial of our business plan and I hope you enjoyed the presentation. \>\> \[applause\]