The Claude Code playbook for large engineering teams | Brian Scanlan (Intercom)
TL;DR
Intercom doubled R&D throughput in 9 months — Brian Scanlan says merged PRs per R&D head are up 2x after going “all in” on Claude Code, with the company now above 90% AI-generated code and asking why it can’t reach 10x.
The real unlock wasn’t autocomplete — it was imagination becoming the bottleneck — Brian points to the Nov/Dec inflection and the post-Christmas Claude Code wave as the moment AI stopped needing heavy “massaging” and started turning ideas directly into working software.
They treated internal AI adoption like a product, with telemetry everywhere — Intercom wired Claude Code usage into Honeycomb, Snowflake, and S3, tracks skill invocation and session outcomes, and even built personalized feedback tools so employees aren’t just handed an API key and told “good luck.”
Quality didn’t collapse; their Stanford analysis suggested it improved — Alongside faster shipping, Brian says time from first line of code to public release is dropping, incident rates aren’t rising, and outside research on Intercom’s data indicated code quality was getting better, not worse.
Their leverage comes from opinionated guardrails, not just raw model access — Intercom built hooks and skills like a mandatory
create_prflow, flaky test fixers, CI integrations, and internal admin tools so agents behave more like the company’s best engineers inside a 15-year-old Ruby monolith.Scanlan’s big claim is that backlog zero is now realistic — He argues teams can finally speedrun tech debt, flaky tests, CI bottlenecks, and old architectural regrets, turning work that once required roadmap lobbying into single Claude Code sessions.
The Breakdown
The moment Intercom decided AI wasn’t optional
Claire opens by framing Intercom as one of the rare companies that “met the moment” twice: first in product, then in how the org itself works. Brian says that made engineering adoption easier because leadership had already decided AI would reshape the company; the frustration was never whether it mattered, but why the big productivity jump wasn’t happening faster.
When Claude Code stopped feeling like autocomplete
Brian says the big inflection came around November/December, then really hit after Christmas, when people came back from break convinced everything had changed. His line is the one that sticks: your “imagination is now the barrier, not the tool,” because the work shifted from coaxing the model to just giving it ideas and seeing what happened.
The chart: 2x more PR throughput, and now CI is melting
Intercom’s CTO set a blunt target: double R&D throughput, measured crudely but usefully through pull requests. Brian shows merged PRs per R&D head doubling over nine months, says their CI system “melted” under the load and got 10x more expensive, and notes the bottleneck has now moved from build infrastructure to code review — which is exactly the kind of problem you get when the machine is actually working.
Why they believe speed is also improving quality
Brian pushes back hard on the idea that more PRs means more slop. Intercom tracks time from first line of code to shipping, sees more real features getting out, shares data with a Stanford research group, and so far isn’t seeing more incidents; the Stanford readout he mentions even suggested code quality was improving.
A lobster emoji demo inside a million-line Ruby monolith
To make this concrete, Brian uses Claude Code to add a lobster-emoji Rails redirect inside Intercom’s ancient “majestic Ruby on Rails monolith,” a codebase older than the company itself. The demo is trivial on purpose: his point is that even boring tasks now move from thought to production absurdly fast, and that’s exactly why the org can reserve human attention for harder problems.
Skills, hooks, and the software factory mindset
One of the best bits is Intercom’s obsession with PR descriptions. Brian says Claude initially wrote terrible ones that just paraphrased code, so they built a create_pr skill, measured quality with an LLM judge, and then enforced the skill with a hook that blocks GitHub PR creation unless the better flow is used — a tiny change, but emblematic of how they’re trying to build an “IKEA factory” for software.
Telemetry for humans, not just agents
Intercom logs skill usage to Honeycomb and stores anonymized session data in S3 so they can study how people actually use Claude Code across the org. Brian shows an internal dashboard that grades his own behavior, including a note that he was being unproductive by arguing with Claude instead of fixing his memory files, which captures the vibe perfectly: the company is instrumenting its AI rollout like any other serious product.
Flaky tests, backlog zero, and the case for going all in
Brian’s favorite example is a flaky-spec skill that evolved from “help me fix this test” into a self-improving workflow that updates itself, fans out to similar failures, and now works at what he calls a distinguished-engineer level. That leads to his broader claim: backlog zero is realistic now, because all the internal work teams used to beg roadmap space for — CI fixes, tech debt, architecture cleanup — can be speedrun instead of endlessly deferred.
The next frontier: products built for agents, not just people
Near the end, Brian shifts from internal engineering to customer experience and argues SaaS products need to become agent-friendly. His worry is that agents default to “build it yourself” unless vendors provide better CLIs, APIs, MCPs, and onboarding flows, so Intercom is exploring ways to let an agent sign up, verify email, and configure Finn without bouncing the human back into every annoying step.
Why the culture got better, not worse
The emotional throughline is that this has made work more fun. Brian says the last three months have been the most enjoyable of his career because he can finally do the things he always wanted to do, ship customer-facing features in hours, and spend more time giving others permission to try ambitious AI workflows — with a simple leadership rule: if something goes wrong, blame me.