Everyone has better models. Why do some teams fly while others don't?
Models got dramatically better over the past year. But they got better for everyone. In that time, our team went from 5.3 to 13.7 PRs per person per month, and I personally went from 4.8 to 51.3. Did you improve that much too? If not, here are a few things that work for us and might well help you as well. Same model, different results Claude, GPT, Gemini. We all have access to the same models. So…
Over the past year, artificial intelligence models have significantly improved, leading to increased productivity for all teams. However, the performance of individual teams varies. In this account, a team's progress is detailed, along with the strategies that contributed to their success. The team's utilization of the Keboola Connection repository played a crucial role in their accelerated progress.
As of December 2024 to May 2025, the team recorded an average of 5.3 PRs per person per month, while the individual's personal PR count was 4.8. By June to October 2026, the team's average rose to 13.7 PRs per person per month, and the individual's personal PR count increased dramatically to 51.3. The team's median time to merge decreased from 48.9 hours to 18.6 hours.
Several factors contributed to the team's improved performance. To begin with, the team leveraged the same AI models as their competitors. The team implemented a set of rules in a single document called AGENTS.md, which served as a single entry point for both humans and AI agents. This streamlined approach facilitated easier collaboration between humans and AI.
A key component of the team's success was the development of the ./bin/kbc CLI, which provided a single command for various tasks, including running tests, static analysis, and preparing the environment. The CLI had 65 tasks and 21 of the team's 33 skills referenced it. Additionally, the team created a skill called kbc doctor, which verified the environment's readiness. Each team member used the CLI frequently, and even the AI agents utilized it multiple times.
The team also developed skills for repetitive work, such as the migrate-endpoint skill, which handled endpoint migrations. The skill significantly reduced the migration workload, with the team migrating 23.2 endpoints per month compared to 3.7 previously. The migrate-highway skill further streamlined migrations, resulting in only 72 out of 270 old Zend actions remaining.
Another essential skill was the run-e2e-ci skill, which triggered the E2E Single Suite workflow in the Continuous Integration (CI) pipeline. This skill ran 238 times in a single month, freeing team members from setting up local environments and testing complex architectures such as Snowflake and BigQuery.
The babysit-pr skill was crucial in managing PRs, handling failing CI and review comments, and assisting AI agents in providing ready-to-review PRs. The ai-review label added a second review bot to handle cases when Copilot reached its limits. The team's productivity improved, with the median time to merge dropping from 49 hours to 26 hours.
Lastly, the ship-task skill connected Linear tasks to PR creation, reducing the median time to merge for PRs created via this skill from 49.2 hours to 15.9 hours. The agent handled task analysis, implementation, review, and PR creation, ultimately resulting in a more efficient workflow and faster development cycles.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
