How I AIHow Coinbase scaled AI to 1,000+ engineers | Chintan Turakhia
Claire Vo and Chintan Turakhia on coinbase’s playbook for scaling AI adoption across 1,000+ engineers.
In this episode of How I AI, featuring Claire Vo and Chintan Turakhia, How Coinbase scaled AI to 1,000+ engineers | Chintan Turakhia explores coinbase’s playbook for scaling AI adoption across 1,000+ engineers Chintan Turakhia explains how Coinbase moved from superficial AI trials ("hello world" usage that didn’t stick) to sustained adoption across 1,000+ engineers by treating AI as an “adapt or die” accelerant rather than a mandate.
At a glance
WHAT IT’S REALLY ABOUT
Coinbase’s playbook for scaling AI adoption across 1,000+ engineers
- Chintan Turakhia explains how Coinbase moved from superficial AI trials ("hello world" usage that didn’t stick) to sustained adoption across 1,000+ engineers by treating AI as an “adapt or die” accelerant rather than a mandate.
- The approach centered on a high-conviction, hands-on leader demonstrating real wins, focusing first on eliminating engineering toil (tests, linting, PR setup) and creating social proof via shared channels and live “PR speed runs.”
- He emphasizes measuring impact through end-to-end cycle time—ticket to production/user value—then compressing each stage (draft PR creation, review time, release) to unlock customer feedback loops.
- The episode also demos practical systems: using Cursor analytics to cohort users and generate a playbook, and an internal Slack/Linear agent (“Claude Bot”) that turns live feedback into tickets and PRs while meeting security requirements.
IDEAS WORTH REMEMBERING
5 ideasAI adoption fails when it’s trialed, not operationalized.
Coinbase saw early Copilot/tool adoption spikes that faded because engineers tried it once, found it lacking, and wrote it off. The fix was persistent, daily use until repeatable workflows emerged.
A single credible, hands-on champion can change the culture faster than policy.
Chintan argues leaders must “show, not tell” by using the tools in real coding work, learning failure modes, and demonstrating concrete wins—engineers ignore decrees but follow evidence.
Start with soul-sucking toil to create immediate trust and pull demand.
Targeting unit tests, linting, and other “paper cut” tasks made AI valuable quickly and freed engineers to do higher-leverage work, building momentum for broader use.
Create viral visibility of wins (and losses) inside existing communication hubs.
A dedicated channel (“cursor-wins”) let engineers broadcast successes, prompting peers to copy techniques. Keeping the magic in Slack makes it observable and shareable, driving organic spread.
Time-boxed “PR speed runs” convert skepticism into belief in minutes.
By having everyone ship a trivial PR during an all-hands, teams experienced a rapid, tangible output spike (e.g., 70 PRs in 15 minutes; later 300–400 PRs in 30 minutes company-wide), making velocity real and undeniable.
WORDS WORTH SAVING
5 quotesIt’s not only possible, it’s adapt or die.
— Chintan Turakhia
Show the engineers, not just tell. And the worst thing any eng leader could do is just be like, 'I decree you must use AI.'
— Chintan Turakhia
No one’s getting bonus points for memorizing Git commands.
— Claire Vo
It was really sort of a death to status updates, long live building moment.
— Chintan Turakhia
My calendar’s empty… the coordination overhead… No, you just do things.
— Chintan Turakhia
QUESTIONS ANSWERED IN THIS EPISODE
5 questionsIn your early “trough of sorrow,” what specific prompting/workflow changes turned Cursor from “kinda sucked” into daily value for you personally?
Chintan Turakhia explains how Coinbase moved from superficial AI trials ("hello world" usage that didn’t stick) to sustained adoption across 1,000+ engineers by treating AI as an “adapt or die” accelerant rather than a mandate.
For the PR speed run: what guardrails did you set (scope, review expectations, rollback rules) to prevent chaos while still “breaking the rules”?
The approach centered on a high-conviction, hands-on leader demonstrating real wins, focusing first on eliminating engineering toil (tests, linting, PR setup) and creating social proof via shared channels and live “PR speed runs.”
You reduced PR review cycle time ~10× (150 hours to ~15 hours). What concrete process/tool changes caused the biggest drop?
He emphasizes measuring impact through end-to-end cycle time—ticket to production/user value—then compressing each stage (draft PR creation, review time, release) to unlock customer feedback loops.
How do you prevent “agent-generated PR spam” from overwhelming reviewers and CI/CD, especially after large-scale speed runs?
The episode also demos practical systems: using Cursor analytics to cohort users and generate a playbook, and an internal Slack/Linear agent (“Claude Bot”) that turns live feedback into tickets and PRs while meeting security requirements.
What Cursor rules or templates were the highest leverage for your team (unit tests, linting, PR descriptions, repo conventions)?
Chapter Breakdown
Why AI adoption in big engineering orgs is “adapt or die”
Claire and Chintan frame the core tension: many still doubt that large, mature engineering orgs can meaningfully adopt AI. Chintan argues it’s not only feasible at 1,000+ engineers—it’s becoming existential and a major leverage point for productivity.
The catalyst: rewriting a major product under extreme timelines
Chintan explains the organizational and product pressures that made acceleration urgent: a major rewrite from a self-custody wallet into a consumer social app. The team needed consumer-grade quality while competing with companies that have much larger teams and head starts.
Early tool failures and the “make it stick” adoption problem
The org had seen initial spikes with tools like Copilot, but usage didn’t persist after novelty wore off. Chintan describes how early Cursor/model limitations made it easy for engineers to dismiss AI, creating a trough that leadership had to push through.
Leadership tactic: hands-on conviction + “show, don’t tell”
Both speakers emphasize that AI transformation needs a credible, technical leader using the tools daily. Chintan’s approach is to demonstrate concrete wins rather than mandate usage, and to focus on removing toil that drains builders.
Workflow breakthroughs: Cursor rules, PR automation, and killing busywork
Chintan details small but powerful workflow shifts—like using Cursor rules and having the tool create draft PRs automatically. The big insight: engineers shouldn’t waste time on “memorize Git commands” tasks when AI can handle them reliably.
Social proof at scale: “cursor-wins” channel and community learning loops
Adoption accelerates once engineers see peers succeed and share examples. The team uses a dedicated Slack channel for wins (and optionally failures) to spread tactics, prompts, and rules organically across the org.
The PR Speed Run: a live event that converts skeptics into users
Chintan introduces a high-leverage adoption mechanism: a timed “speed run” where everyone ships a trivial PR using AI. The result was immediate proof of feasibility—70 PRs in 15 minutes on one team, then 300–400 PRs in 30 minutes company-wide—plus valuable infrastructure pressure testing.
Measuring real impact: compressing end-to-end cycle time
Chintan argues against vanity metrics like “AI lines written” and focuses on time-to-value: from ticket/feedback to changes landing for users. He highlights dramatic PR review cycle time reduction (about 10x) and a broader obsession with closing the loop from user feedback to shipped fixes.
Using AI to analyze AI adoption: Cursor analytics → cohorts → playbooks
Chintan demonstrates pulling Cursor admin analytics CSVs and using Cursor itself to cluster users into adoption cohorts (agent-heavy, tab-heavy, balanced, minimal/inactive). The workflow generates scripts, dashboards, and cohort-specific guidance to systematically move people toward power-user behavior.
From messy feedback to shipped code: the “feedback capture → Linear → PR” pipeline
Chintan shows a system that captures live voice feedback (e.g., at a “feedback cafe”), extracts actionable bugs with an LLM, creates Linear tickets, then triggers an internal Slack bot to start implementing fixes. The goal is to eliminate the slow human transcription and triage chain that causes users to churn before issues are addressed.
Why build an internal Slack/Linear agent: security, context, and virality
Chintan explains why Coinbase built an in-house bot (“Claude Bot”) rather than relying solely on external agents: security constraints and the need for deep internal context. Slack is the distribution layer—doing the “magic” in public makes it go viral and drives adoption organically.
Personal AI workflows: calendar automation and reverse-engineering taste
Chintan shares lightweight personal use cases that make AI feel instantly valuable: turning school emails into calendar events and analyzing tasting notes to infer champagne preferences. He extends it to restaurant wine menus to recommend best-fit bottles and values based on learned preferences.
What changes for leaders: fewer meetings, more code, and better prompting
In a lightning round, Chintan describes how AI reduces coordination overhead so dramatically that his calendar is nearly empty and he’s writing far more code. He also shares pragmatic prompting tactics—insisting on corrections, escalating clarity, and even jokingly “threatening” model switching to get better output.
EVERY SPOKEN WORD
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome