Source: raw/How_the_Claude_Code_team_uses_Claude_Code.md — official Claude YouTube channel (@claude), video ID S-sYlFiGFv8 (https://www.youtube.com/watch?v=S-sYlFiGFv8), auto-caption transcript fetched 2026-10-30, about 4,500 words. Publish date not in the transcript.

Speakers: members of the Claude Code team in a roundtable. The transcript names only Robert (building Claude Tag) and Sid (who built workflows); Boris is mentioned. The host and the other speakers are not named.

A team roundtable on how Claude Code’s own engineers work about a year after one of them joined. The through-line: they have moved from approving tool calls one by one to handing Claude a goal, and most of their work now runs through Claude Tag in Slack. They describe how loops, workflows, code review and verification grew out of specific limits they hit, and why they keep deleting harness features the newer models no longer need.

Key Takeaways

  • Most work now runs through Claude Tag. One engineer puts it at “70 to 80% of my work”; the remaining 20% is the TUI or desktop app, to refine something or to “micromanage my Claudes.” The tasks they give Claude Tag are “worlds more complicated” than “implement this class.”
  • Goals, not transcripts. The team has gone from caring about “each individual tool call and each individual decision” to “I have a goal… you just kind of give that goal to the model.”
  • Shelf life has collapsed. Product technology used to last years; with models “it fundamentally shifts underneath you every two months.” You have to stay at the frontier while still giving value to people using today’s models.
  • Harness features are temporary. To-do lists were the first “ride the wave” moment (around Sonnet 3.5, a model given five things would do three and give up; a to-do list fixed it). A year later “you don’t really need the to do list anymore.” Much of the harness exists “to cover for some of the failure modes” of the current model, and gets removed as models improve.
  • AskUserQuestion went the same way. The tool was first placed after planning, then made something Claude can call; it took a long time to get Claude to call it well. Its designer now mostly asks Claude for an artifact instead: an HTML page with diagrams and mockups that asks the questions.
  • The road to loops: laptop → remote hosted dev boxes over SSH → Claude Code on the web (hosted containers) → routines. Example routine: every day, read all incoming feedback, bucket it by importance, and fix the items Claude has high confidence in fixing. One speaker says cloud setup “10x my productivity.”
  • Human code review moves up a level. Claude handles the nitpicks reviewers used to post to prove they read the code. The human’s job is the bigger picture (why the API is structured this way, why service boundaries sit where they do), and Claude can pull that context into the review for the human.
  • Workflows came out of code review. Code review was the team’s first big fan-out: send agents out to find bugs, then run an adversarial review of each bug from three perspectives to check it is real, and return only the important ones. Sid calls it a MapReduce problem; the host calls it test-time compute.
  • Deterministic code builds trust. Because Claude writes code to orchestrate the subagents, a for loop over items “is not going to skip one of the items.” That mix of deterministic code and agentic behavior is what makes the speaker trust a workflow.
  • Claude Tag’s UI is a tool call. When Claude Tag posts in Slack it is calling a messaging tool; its internal monologue is not shown (there is a link to the full transcript). Robert found this “a little scary” at first, then freeing: “Claude is choosing what to say to me and when.”
  • Three primitives behind the work: verification, code review, and feedback (a connection to your data or event store: metrics, Slack, GitHub issues). When Claude opens a Claude Code PR “it tests it, it sends me screenshots.”

Details

From local sessions to routines

The loop story starts with a practical problem: an engineer running agents on a laptop had to shut it down to go home, and the agent stopped. Leaving it open in the parking lot drew looks. The steps that followed:

  1. Remote hosted developer boxes. The agent keeps running, but you have to SSH back in.
  2. Claude Code on the web. Hosted containers that keep running in the background. Setup is the hard part, because the container needs access to your dev environment. “It’s totally worth it.”
  3. Routines. Once Claude runs in the cloud, you can schedule it: the daily feedback-triage routine above. This is the “loop journey that Boris talks about a lot”: breaking the session boundary and prompting something one level above the session.

See the cloud sessions field guide for the current setup.

Code review and the fan-out pattern

  • What Claude takes over: the “three little nit picks” a human reviewer posts to signal they read the code. Claude finds and fixes those on its own.
  • What stays human: context Claude has not internalized, such as why an API or a service boundary is the way it is. Claude’s role is to surface that context and hand it to the reviewer.
  • The fan-out shape: fan out to find candidate bugs → for each bug, adversarial review from three perspectives → filter → return only what needs attention. The same shape works for performance issues and general deep research.
  • Sid’s trip example: planning a trip to Tahoe, run about ten searches (bathhouses, hotels, Airbnb, places to go), then agents that sort for the most interesting results, then agents that verify them, then a summary. “If I read the fan out output I’m going to go crazy,” so the filter step is the point.
  • Claude designs its own harness. “Claude is actually really good at making its own harnesses”: what the fan-out should look like, its topology, how one agent’s output feeds the next level, and the final summary. Shipped as dynamic workflows.

Claude Tag, as its builders use it

  • Robert: Claude Tag builds Claude Tag. His focus is making the development environment and dev loop easy for Claude to use, so it can do everything a human does to build the software and test it end to end. That gets harder as the software gets more integrated.
  • Another engineer’s tool launch, all in Slack: ask who to talk to about an idea → get stakeholders → have Claude make mockups (viewable on a phone) → implementation → add events → deploy internally → Claude Tag monitors usage and tags the engineer when someone leaves feedback. When the funnel looked weak, instead of dictating a fix they asked Claude to “improve the funnel” and come up with ideas, giving their own idea only as an example.
  • Verification without opening the TUI. For a Claude Code change, one speaker first asked Claude to record itself using the TUI, then cloned the branch to try it “just for sanity’s sake,” and expects to stop doing even that.

What the team misses

  • Performance engineering. One speaker used to love digging into a system’s performance; “Claude is much better at that than I am now,” but the speedups still land. Their attention has moved to getting ideas from prototype to production faster.
  • Hand-crafted UI. Another spent a day on a CSS recreation of macOS 10.4 Aqua buttons from stacked radial gradients; Claude can do that now. They compare it to building games in PowerPoint as a child: “the entirety of software engineering” now feels within reach.
  • The closing view: software engineering is “the profession of change”; the problems differ, but it is still problem solving.

Try It

  • Delete a scaffold. List the harness rules in your CLAUDE.md, skills and hooks that exist to work around a model weakness (forced to-do lists, step-by-step checklists). Remove one, rerun a task on the current model, and keep it removed if nothing breaks.
  • Ask for a question artifact. Instead of a long back-and-forth, ask Claude to build an HTML page that lays out the options with diagrams or mockups and asks you the open questions. See HTML Effectiveness (Thariq) for the pattern.
  • Copy the feedback-triage routine. A daily routine: read new feedback, bucket by importance, open fixes only for items it is confident about.
  • Add adversarial review to a fan-out. For any fan-out (bug hunt, research), add a step that reviews each finding from three perspectives before it reaches you. In Claude Code this is what workflows are for.
  • Ask for proof, not a report. Tell Claude to attach screenshots or a recording of itself using the change before you look at the diff. See Verification Loops and the Verify Skill.

Open Questions

  • Speaker names beyond Robert and Sid, and their roles, are not in the transcript. “Sid” is probably Sid Bidasaria, co-author of the workflows best-practices post cited in Dynamic Workflows.
  • The “70 to 80%” and “10x” figures are one engineer’s own estimates, not measured.
  • The video’s publish date is not in the transcript.