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

Key Concepts

• Vibe Coding: A term referring to writing code by relying on AI to generate solutions without necessarily understanding the underlying implementation, which the speaker warns can lead to "slop code."
• Comprehension-First AI: The strategy of using AI primarily to understand complex, legacy codebases rather than relying on it for code generation.
• Agentic Tools: Software agents (like Sentry’s Warden or Junior) that perform autonomous tasks such as code reviews or bug fixes.
• Technical Debt: The implied cost of additional rework caused by choosing an easy solution now instead of a better approach that would take longer.
• Mental Model Alignment: The process of ensuring the developer’s understanding of the system matches the AI’s output before proceeding with implementation.

---

1. Main Topics and Key Points

Priscilla, a Senior Software Engineer at Sentry, discusses her transition from traditional coding to an "agent-orchestrating" workflow.

• The Shift in Productivity: The speaker emphasizes that in a complex, 15-year-old codebase with 100+ PRs merged daily, the primary value of AI is not writing code, but comprehension.
• Data-Driven Usage: After analyzing 116 of her own AI interaction sessions, she found that 67% of her usage was for comprehension, while only 2% was for code generation.
• The "Catch Me Up" Skill: She developed a custom, local AI skill that structures queries into six exploration modes: Architecture, Convention, Feature Trace, Syntax, Testing, and History. This allows her to quickly grasp the context of unfamiliar parts of the Sentry codebase.

2. Real-World Applications at Sentry

Sentry utilizes several internal AI tools to manage their complex observability platform:

• Abacus: Tracks internal AI usage.
• Warden: An automated code review agent integrated into Pull Requests.
• Junior: A Slack-integrated bot that analyzes bug reports or UI complaints, creates a corresponding PR, and fixes the issue.
• AI SDK Testing Repository: A dedicated space for testing AI integrations, where the speaker was instructed to "only prompt" until the desired result was achieved.

3. Methodologies and Frameworks

The speaker advocates for a structured approach to AI-assisted development, building upon the "Research, Planning, and Implementation" framework:

1. Research (Comprehension): Use AI to understand the existing architecture and history.
2. Alignment: Verify the AI’s research against your own mental model to ensure it is on the right path.
3. Planning: Instruct the AI to create a plan based on the verified research.
4. Implementation: Execute the code generation only after the previous steps are validated.

4. Key Arguments and Perspectives

• Quality Over Speed: The speaker warns against "vibe coding," which can lead to "slop code." She argues that developers must maintain ownership of the code they ship, as it is their responsibility to ensure the codebase remains maintainable.
• The "Cheapest Senior Engineer": AI is framed as an tireless teammate that can answer any question, regardless of how basic, making it an invaluable tool for onboarding and navigating legacy systems.
• The Danger of Disconnection: Citing Armin Ronacher (creator of Flask), she warns that if developers no longer understand the code in their own repositories, the industry is heading toward a "disaster."

5. Notable Quotes

• "The biggest unlock from AI in a large codebase isn't generation. It's comprehension."
• "Don't ship slop code into the code base that pays your salary. Ship keynote code."
• "AI is the teammate who never gets tired of your questions. So, there is no dumb question."

6. Synthesis and Conclusion

The core takeaway is that AI should be treated as a context-provider rather than a code-writer. By shifting the focus from generation to comprehension, developers can navigate massive, complex codebases more effectively while maintaining high standards of code quality. The speaker encourages developers to track their own AI usage patterns to identify where they can improve their workflow, ultimately advocating for a "human-in-the-loop" approach where the developer remains the architect and the AI acts as the highly efficient, knowledgeable assistant.