How One CEO Shipped 52 PRs While on Vacation

Theo, CEO of t3.gg, experienced a significant increase in his weekly pull request (PR) contributions between mid-July and August. His output escalated from an average of 1-3 PRs per week to an astonishing 23, then 37, 41, and eventually 52 PRs in a single week. This impressive growth occurred while he was traveling for Defcon, indicating that his new workflows enabled him to maintain high productivity even outside of a typical office environment.

Theo emphasized that he has developed productivity workflows that enable him to be as effective at coding as he desires, despite also running multiple companies and managing personal life responsibilities. He explicitly challenges the notion that shipping code effectively necessitates being a full-time engineer. By implementing these new flows, he demonstrates that it's possible to rival the output of other high-volume contributors while balancing diverse commitments.

A viewer-reported issue brought to light a significant limitation in the dollar-sign trigger syntax for AI skills. Theo confirmed that if a skill is configured to be invoked only manually, it will exclusively work with the slash command and will not function with the dollar-sign command version. This discrepancy between AI-triggered and manual-only skills has been identified as a confirmed bug, highlighting a gap in the current implementation.

Theo underscored the critical importance of offloading heavy coding tasks, especially those involving AI agents, to remote infrastructure rather than running them on one's primary laptop. He stressed that the main machine used for other activities should not be bottlenecked or required for these intensive coding tasks to be completed efficiently. His recommendation is to utilize a remote Linux box to prevent hardware constraints from impeding progress, and to leverage work trees for parallel task execution, ensuring that local development remains unhindered.

Theo explained that providing AI agents with context about user impact significantly improves the reliability of their coding fixes. By explicitly stating how many users are affected by an issue, the AI gains a clearer understanding of the problem's importance, leading to more reliable and robust solutions. Furthermore, clearly defining priorities, such as favoring simplicity over speed, ensures that the AI generates code of higher quality that aligns with specific project goals. This contextual information transforms a purely experimental task into one that demands dependable execution.

Theo outlined his method for delegating the triaging of pull requests (PRs) to AI sub-agents. He uses multiple threads, each employing different AI models, to audit open PRs in his repository, which he had lost track of. The AI agents are tasked with going through all PRs and prioritizing them based on the ease of action required. This cost-effective workflow, utilizing fast and inexpensive models, helps him efficiently manage a large volume of contributions.

Theo highlighted the strategic use of AI tokens to reduce stress and the likelihood of decision fatigue or distractions, particularly for individuals with ADHD. By leveraging AI to simplify complex technical explanations, he significantly reduces the mental load associated with understanding intricate problems. This approach helps maintain productivity during long tasks, preventing the temptation to switch to other work.

Theo has optimized his workflow to make the creation of new threads for issues virtually frictionless, significantly boosting his pull request (PR) throughput. This system encourages immediate intervention whenever an issue is noticed, allowing him to close issues exceptionally fast. By forcing himself to address detected bugs immediately, he maintains a continuous, low-effort engagement with the repository's health, ensuring problems are tackled as soon as they arise.

Theo identified a user experience (UX) issue where the AI model incorrectly applied only the front-most skills in certain CLI commands. This limitation led to user confusion because other intended skills were not executed. To prevent this, Theo suggested implementing UI indicators for non-frontable commands, which would clearly communicate to users when a command cannot be fully applied, thereby improving clarity and reducing frustration.

Theo introduced a manual linking mechanism to attach a pull request (PR) to an existing conversation thread. This functionality is crucial for enabling auto-settlement, a feature that automatically closes threads once the associated PR is merged. This streamlines the workflow by ensuring that once a code change is integrated, the related discussions are automatically resolved, reducing manual cleanup and improving overall project organization.

Theo demonstrated the critical need for manual oversight by rejecting an unreliable recommendation from the Luna AI model. He chose to archive Luna's output due to its poor quality, emphasizing that while AI is helpful, human judgment remains essential for tasks involving complex native changes. This highlights that model suggestions, especially for intricate code modifications, require careful scrutiny to ensure integrity and prevent the introduction of new issues.

Theo successfully leveraged AI agents to identify and draft professional communication regarding a bug. This effort led to external confirmation of the issue by Boris from the Claude Code team, who commented in the thread and verified that it was a genuine bug. This collaboration underscores the effectiveness of using AI not only for problem identification but also for facilitating clear and concise communication with external teams.

Theo invested considerable time and resources into integrating a built-in Tailscale share dev command to facilitate remote code testing. This command allows him to test code changes from anywhere, including locations with unreliable internet such as hotel Wi-Fi. Additionally, he completely overhauled the bundling process, significantly improving performance and ensuring a smoother testing experience even under challenging network conditions.

Theo implemented a clever fix to automate preview builds for macOS, which can now be triggered simply by adding a specific label. This innovation eliminates the need for manual local builds, significantly reducing friction in the development workflow. Furthermore, he removed the GitHub sign-in requirement for these builds, making it easier to test code on secondary machines without additional authentication steps. These improvements streamline the testing process, allowing developers to focus more on coding and less on setup.

Theo articulated a core philosophy centered on identifying and eliminating friction points within development workflows. He urged developers to pay close attention to what prevents them from merging code, as these obstacles represent opportunities for significant improvement. By continuously making small, incremental enhancements to projects, workflows, and tools, developers can achieve exponential increases in throughput. The ultimate goal is to cultivate a sense of confidence when hitting the merge button, knowing that the process is robust and streamlined.

Theo advocates for a strategy to "de-risk the merge button," where developers should not even review a pull request (PR) until multiple AI agents have expressed high confidence in its readiness. This approach emphasizes creating robust safety nets through automation to prevent breaking changes from reaching a broad user base. The focus is not on total prevention of issues but on quickly catching and addressing them, ensuring that the merge process is as reliable and stress-free as possible. By building these layers of automated validation, developers can merge code with greater assurance.