ClaudeHow the Claude Code team uses Claude Code
At a glance
WHAT IT’S REALLY ABOUT
How Claude Code engineers delegate goals via Slack-native Claude Tag agents
- The Claude Code team increasingly does 70–80% of their work through Claude Tag in Slack, delegating complex goals rather than supervising token-by-token transcripts.
- They build on rapidly shifting model capabilities, treating many features as temporary scaffolding that can be removed as models improve.
- They evolved from local, laptop-bound agents to cloud-hosted containers that enable persistent loops, routines, and ongoing background work.
- Their code review approach uses Claude to remove low-value nitpicks and apply fan-out, adversarial validation, and coalescing (“test-time compute”) to surface the most important issues.
- They emphasize verification and feedback loops—tests, screenshots, event data, and Slack feedback—to ship faster while maintaining confidence.
IDEAS WORTH REMEMBERING
5 ideasThey’ve moved from micromanaging tool calls to delegating outcomes.
The team now works from a “goal-first” posture—handing Claude Tag an outcome to achieve—rather than supervising every tool call or token-level decision in a transcript. This shift is enabled by stronger models and a Slack-native interface that can pull rich product context from team discussions.
AI-product features have a short shelf life; build with detachment and replaceability.
Because model capabilities change dramatically on ~two-month cycles, features that once felt essential (like explicit to-do lists for long-horizon work) can become obsolete quickly. The team treats many product mechanisms as temporary scaffolding for current model failure modes.
Slack-native agents are powerful because they inherit organizational context.
Claude Tag works inside Slack, which lets the agent access surrounding context—product decisions, stakeholder discussions, and ongoing feedback—improving its decision-making beyond what’s in a single coding session. This context access also changes how work is coordinated (e.g., mobile-friendly collaboration, faster stakeholder alignment).
Cloud-hosted loops turn agents from sessions into ongoing processes.
A key evolution is moving agents off the laptop and into hosted environments that can run continuously. Once agents run in the cloud, the team can schedule routines (e.g., daily triage of feedback) and run persistent “loops” that break out of one-off chat sessions.
Code review shifts to high-level intent, powered by fan-out and verification.
Claude helps humans avoid performative review by autonomously handling many nitpicks and by surfacing the highest-leverage review questions (architecture, API boundaries, intent). The team also uses fan-out + adversarial perspectives + coalescing (“test-time compute”) to find and validate issues at scale.
WORDS WORTH SAVING
5 quotesI think we've gone from just, you know, caring about the transcript so much, and, like, caring about each individual tool call and each individual decision that the model makes to just zoomed-out view of like, "I have a goal, and I'd like to achieve this goal," um, and we-- and you just kind of give that goal to the model to, to achieve for you.
— Sid Bidasaria
With models and the AI models, it's like the technology fundamentally shifts underneath you every two months.
— Sid Bidasaria
Like, you gotta be very unattached to the things you're building-
— Sid Bidasaria
I think like a common thing that I saw happen in traditional, like human code review- ... is you would have someone go and review the, the code, and they would like pick out like, you know, three little nitpicks to post, and it was kind of just like I'm signaling to you that- ... I read this code.
— Robert Boyce
Um, and like being forced even- ... to let Claude cook.
— Robert Boyce
High quality AI-generated summary created from speaker-labeled transcript.
