How the Claude Code team uses Claude Code

From a coding assistant to a delegated worker

The latest account of how Anthropic’s Claude Code team uses its own product shows a sharp change in the unit of software work: instead of asking a model to implement a class, patch a function or explain a file, team members increasingly give Claude a goal and let it choose the path [1]. That shift matters because Claude Code began as a developer-facing coding agent, but the team’s own habits now look closer to managing a set of cloud workers than supervising a single assistant in a terminal [1].

The central interface in the new workflow is Claude Tag, the Slack-native surface layered on top of Claude Code [1]. In the reported discussion, team members said that Claude Tag has become the place where much of their daily work happens, with one estimate putting 70% to 80% of that work inside the Slack-based agent rather than in the terminal or desktop app [1]. A separate current write-up of the same video frames the change similarly: a year ago the flow was prompt, inspect, approve and repeat; now the team hands Claude Tag a higher-level goal and lets it continue through more of the execution chain [2].

This is not just a user-interface preference. Slack gives the agent access to the place where product decisions, feedback threads, bug reports and team priorities are already being discussed [1]. That ambient context lets Claude make more product-aware judgments than it could if it were limited to a local repository and a fresh prompt [1]. In practice, the agent is being asked not only to write code, but also to understand why a change matters, who should review it, what evidence should be gathered and when a human needs to be brought back into the loop [1].

The product is being rebuilt around model speed

The most important backdrop is the pace of model improvement. Team members described building on technology that can change materially every two months, a cadence that makes many traditional product assumptions unstable [1]. In ordinary software, a workflow feature might remain useful for years; in a frontier-model product, it may be a temporary scaffold for a limitation that disappears in the next model cycle [1].

That creates a different discipline for product management. The Claude Code team has to decide whether a feature solves a durable user problem or merely compensates for a current model weakness [1]. The reported example is the to-do list pattern: earlier models benefited from explicit task lists because they could lose track of long-horizon work, but as memory and coherence improved, that scaffolding became less central [1]. The same pattern appeared with the AskUserQuestion tool, which once helped the agent pause and ask for clarification, but is now being complemented or displaced by richer HTML artifacts, diagrams and mockups that can gather human feedback more effectively [1].

The implication is that dogfooding is not merely a cultural slogan for this team. Because the Claude Code group depends on Claude Code and Claude Tag to build the product itself, obsolete friction becomes visible immediately [1]. If a feature exists mainly because the previous model needed help, the next model can turn that feature into dead weight [1]. That is why the team’s current practice includes deleting or rethinking product surfaces as aggressively as it adds them [1].

Cloud execution changes what can be delegated

The move from local execution to cloud execution is another major break from the earlier Claude Code workflow. When the agent ran mainly on a laptop, its work stopped when the machine closed or the developer walked away [1]. The team’s workflow then moved through remote development boxes and browser-based environments toward hosted containers and persistent execution [1].

That infrastructure change enables a different class of tasks. Instead of asking Claude to complete one local edit session, the team can set up routines that run in the background, triage feedback, cluster related reports, monitor outcomes or fix high-confidence issues automatically [1]. The current video listing also identifies “running loops and routines remotely” as a distinct part of the team’s workflow, placing remote execution alongside auto mode, workflows and other primitives [2].

This is where Claude Code starts to look less like a coding tool and more like an orchestration layer. The agent can be given a recurring job, a stop condition, access to tools and a feedback channel. The human role becomes defining the objective, choosing the guardrails and deciding which outputs deserve trust. That is a profound change from line-by-line prompting, but it is not the same as removing humans. It moves the human checkpoint to the level of goals, verification and exception handling [1].

Code review moves from nits to judgment

Code review is one of the clearest examples of the changing boundary. The old ritual of reviewing a pull request line by line and leaving several minor comments is less valuable when Claude can already catch style issues, run tests, add missing checks and fix many small defects [1]. In the team’s current account, human review is shifting toward architecture, service boundaries, API design and product intent [1].

That shift does not mean code review disappears. It means that the scarce human contribution is no longer proving that someone read the diff. It is deciding whether the change is the right change [1]. Claude can fan out across the codebase, search for likely bugs, compare different hypotheses and surface the strongest findings [1]. Humans then spend more of their attention on context the model may not fully own: why a boundary exists, what a customer promise implies, which risk is acceptable and whether a proposed abstraction will age well [1].

The same fan-out pattern is becoming a broader design motif. Rather than asking one agent to produce one answer, the system can launch multiple lines of investigation, have them critique one another, filter the results and return a smaller set of high-confidence findings [1]. That is especially important in software engineering because correctness is not only a matter of generating plausible code; it depends on tests, adversarial review, system context and operational evidence [1].

Building Claude Tag with Claude Tag

Perhaps the most revealing detail is recursive: the team is using Claude Tag to build Claude Tag [1]. In one reported example, the workflow moved from an idea to stakeholder discovery, mockups, implementation, deployment, event tracking and feedback monitoring largely through Slack [1]. Claude also monitored feedback and tagged the developer when new user signals arrived [1].

That loop explains why the story is about more than productivity. The team is turning its product into part of its product-development nervous system. The same channels that contain feedback, design debate and operational signals are also becoming the place where work is assigned, executed and verified [1]. The result is a tighter cycle between user signal and product change, but also a new dependency: the quality of the team’s engineering process increasingly depends on the quality of the agentic loops it designs.

There is a cultural trade-off. Team members said they miss parts of older software engineering, including the direct craft of manual implementation and the satisfaction of solving low-level problems themselves [2]. But their current description points to a new craft: choosing the right goal, shaping the context, designing verification, knowing when to trust automation and knowing when to intervene [1].

Why this matters beyond Anthropic

The Claude Code team is an extreme case because it builds the agent it uses. Still, its workflow offers a preview of where AI-native engineering may be heading. The work is not simply “AI writes code.” It is a stack of delegation surfaces, persistent execution, contextual memory, automated verification and human judgment [1].

The near-term lesson for other engineering organizations is not to copy every Anthropic-specific tool. It is to notice the direction of travel. Work that is well-scoped, testable, repetitive or feedback-driven is moving into agent loops. Work that requires taste, risk judgment, architecture and product intent remains human-led, but increasingly with agents gathering evidence and executing the chosen path [1].

The deeper lesson is that software processes themselves now have shorter half-lives. If the model underneath a product changes every few months, teams cannot treat today’s prompt patterns, review rituals or interface features as permanent [1]. The Claude Code team’s own usage suggests that the most durable skill is not memorizing a particular workflow. It is building an organization that can keep redesigning the workflow as the agent becomes more capable [1].

Sources from the last 72 hours

1. [1]How the Claude Code team uses Claude Code · Anthropic · 8news.aiSep 2, 2026, 7:07 PM UTC
2. [2]How the Claude Code team uses Claude CodeSep 3, 2026, 1:26 AM UTC
3. [3]Claude — YouTube Stats, Subscribers & RevenueSep 3, 2026, 12:00 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.