The AI Skill I Rely On Daily — Priscila Andre de Oliveira, Sentry

Priscila Andre de Oliveira, senior software engineer at Sentry, presents a counterintuitive finding from her own AI usage data: comprehension, not code generation, is the dominant value AI delivers in large, mature codebases. She frames herself as an "agent manager" rather than a coder, noting that since December 2025 she no longer writes code directly — she prompts exclusively, with Claude as her primary collaborator on features, bug fixes, and cross-repository refactors.

Sentry's engineering context makes the stakes concrete: the codebase is 15+ years old, has roughly 100 PRs merged every single day across 400 employees and open-source contributors, serves 100,000 organizations, and is continuously evolving with component deprecations, new lint rules, and architectural changes. In this environment, keeping a mental model synchronized with the codebase is itself a full-time challenge — Oliveira recounts returning from vacation to find her PR buried in conflicts requiring deep contextual catch-up.

To quantify her own AI usage patterns, she asked Claude to analyze 116 of her own chat sessions. The result: 67% were comprehension tasks, 2% were code generation, with the remainder spread across modification, process, review, and other categories. This data point — far more comprehension than generation — drives the entire talk.

From this insight, she built a personal Claude Code skill called "catch me up," stored locally as a Markdown prompt file. The skill structures comprehension queries across six exploration modes: architecture, conventions, feature trace, syntax, testing, and history. During a live demo, she shows the skill applied to Sentry's AI SDK testing repository — a new codebase she was onboarded to with the constraint that she should not write any code herself, only prompt. She queries: "I am a new contributor. Catch me up on how this repository works and clarify what it simulates a Sentry envelope and intercepts during tests." The skill, running on Claude Opus, returns a structured organogram-style summary that answers the envelope question precisely (it intercepts real envelopes via a spawned collector, rather than simulating them), giving her the architectural grounding needed to contribute safely.

Olivieira also uses the skill during PR reviews when she has partial but insufficient context to approve a colleague's change — the skill rapidly surfaces the surrounding architecture so she can make an informed judgment.

On workflow philosophy, she references Jack Nation's blog post on "vibe coding our way to disaster" (drawing on Rich Hickey's simple-made-easy framework), which proposes a three-phase research-planning-implementation loop. She argues a missing fourth step sits between research and planning: the engineer must personally understand what the agent found before authorizing it to plan and implement. Misaligned mental models, she contends, are the root cause of "slop" code. She also cites Armin Ronacher (Flask creator and former Sentry engineer) who wrote: "When more and more people tell me they no longer know what code is in their own codebase, I feel like something is very wrong here."

Sentry's broader AI investment is also highlighted: an internal hackathon produced Abacus (AI usage tracking), Warden (AI code review agent integrated with PRs), and Junior (a Slack bot that ingests bug threads and opens PRs autonomously). Sentry also ran a "quality quarter" to eliminate technical debt — removing TypeScript any types, stale to-dos, and unused feature flags — demonstrating that AI adoption and code quality are treated as complementary, not opposing, priorities.

Key takeaway: in large production codebases, the highest-leverage AI skill is not generating code but using AI as an always-available, infinitely patient senior engineer for comprehension — understanding architecture, tracing features, decoding conventions, and reconstructing history.