How Anthropic Teams Use Claude Code: Security to Legal, 2027
How Anthropic uses Claude Code across Security, Marketing, Design and Legal, with verified results and 30-day plans to roll it out beyond engineering.
Anthropic has published a first-party account of how its own teams use Claude Code. The teams include Security Engineering and Legal, and the examples go well beyond writing software. This guide breaks down what each team did, what inputs they gave the agent and what they reported. It also turns each pattern into a plan you can run this month. At BuildBetter, we work with B2B product teams that are rolling out coding agents across their companies. The question we hear most is not “which model?” but “what context do we feed it?” Anthropic’s examples answer that question better than most vendor marketing does, so we built this guide around them.
The Story: Anthropic Showed How Its Own Teams Use Claude Code
On July 24, 2025, Anthropic published “How Anthropic teams use Claude Code,” a first-party account of how its own teams use the agentic coding tool day to day. The teams range from Product Engineering and Security Engineering to Growth Marketing, Product Design and Legal.
Anthropic reports two headline numbers:
- Inference: research time fell from about an hour to 10–20 minutes, which Anthropic describes as an 80% reduction.
- Security Engineering: incident problems that typically took 10–15 minutes of manual scanning now resolve 3x as quickly.
The post’s core thesis is one sentence: “Claude Code works best when you focus on the human workflows that it can augment.”
Sourcing note: the original anthropic.com/news URL now 308-redirects to claude.com/blog. We checked every fact in this guide against the claude.com version.
The post still matters in 2027 because few first-party accounts describe an AI coding agent being used outside engineering. Any company planning a cross-team rollout can use it as a reference model.
What Anthropic’s Teams Actually Did: The Mechanics
Every team in Anthropic’s post handed Claude Code a specific artifact and asked it to do a specific job. The table below lists only facts reported in the post. Results are Anthropic’s own figures.
| Team | Input given to Claude Code | Task | Reported result |
|---|---|---|---|
| Product Engineering | The codebase | “First stop” for programming tasks, such as finding which files to look at for bug fixes and features | Default starting point for programming work |
| Security Engineering | Stack traces and documentation | Diagnose incident problems | 3x faster than the 10–15 minutes of manual scanning these problems usually took |
| Inference | Model-specific code | Explain functions to team members without ML backgrounds | Research time cut from ~1 hour to 10–20 minutes (80% reduction) |
| Data Infrastructure | Dashboard screenshots during a Kubernetes outage | Diagnose the outage | Identified pod IP address exhaustion, supplied commands for a new IP pool, saved ~20 minutes |
| Infrastructure data scientists | Codebase plus CLAUDE.md files | Help new hires understand data pipeline dependencies | Faster onboarding (no figure given) |
| Growth Marketing | CSVs of hundreds of ads with performance data | Flag underperformers and generate variations using two specialized sub-agents, plus a Figma plugin | Up to 100 ad variations per plugin run |
| Product Design | Figma files | Autonomous iteration loops, plus mapping error states and logic flows to find edge cases | Edge cases found earlier in design |
| Legal | Knowledge of request types and who handles them | Prototype “phone tree” systems to route staff to the right lawyer | Working prototype |
The pattern to notice: every example starts from rich context. That context is a codebase, CLAUDE.md files, stack traces, documentation, screenshots, CSVs or Figma files. In each case, the agent replaced manual scanning of material a human would otherwise have read line by line. It did not invent answers from nothing.
The Questions Anthropic Had to Answer
Read closely, the post answers five questions every company faces when a coding agent spreads past the engineering org.
Which teams outside engineering get value from a coding agent?
Growth Marketing, Product Design and Legal all built working tools or workflows. None of these teams ships production code as its main job. Each still found a repetitive task the agent could take over.
How should new hires use AI to learn a codebase?
New Infrastructure data scientists feed Claude Code the codebase plus CLAUDE.md files to map pipeline dependencies. The CLAUDE.md file gives the agent the context a senior teammate would otherwise explain in person.
How do we use AI during incidents?
Security Engineering fed in stack traces and documentation. Data Infrastructure fed in dashboard screenshots. In both cases the agent did the scanning and proposed a diagnosis.
When do we let the agent run autonomously versus check in?
Product Design runs autonomous loops on Figma files. The incident examples are diagnostic: Claude suggested commands, and the account implies a human ran them. This split is our reading of the examples, not a stated Anthropic policy.
How do non-engineers build their own tools?
Growth Marketing built a sub-agent workflow and a Figma plugin. Legal built a phone-tree prototype. In both cases the people who owned the workflow built the tool.
How Does Anthropic Use Claude Code Internally?
Anthropic uses Claude Code across engineering and non-engineering teams. Uses include:
- navigating codebases
- debugging incidents
- explaining unfamiliar code
- onboarding new hires
- generating ad variations
- finding design edge cases
- prototyping internal tools
How Anthropic handled it: Product Engineering treats Claude Code as the “first stop” for programming tasks. Before an engineer opens a file tree or greps for a function, they ask the agent where to look. Other teams adapted the tool to workflows they already had. The post describes no single prescribed process, which suggests adoption grew from the bottom up, team by team.
Do it this month:
- Pick one recurring task per team where people currently search or scan manually. Examples include finding the right file, reading a log or skimming a spec.
- Make Claude Code the default first step for that task for two weeks. A default changes behavior. An optional tool usually does not.
- Log time before and after. You want your own numbers, not borrowed ones. Anthropic’s 80% figure comes from Anthropic’s staff, codebase and workflows.
- Share wins in a single internal doc so teams can copy each other’s patterns. A good prompt written by Security often transfers to Infrastructure with small edits.
Expect wins to cluster around tasks that are mostly reading and less around tasks that need judgment calls. That distribution tells you where to expand next.
What Are the Best Claude Code Use Cases for Non-Engineers?
At Anthropic, non-engineers used Claude Code to build their own tools. Marketing built ad-analysis and variation workflows. Design mapped edge cases from Figma files. Legal built routing prototypes, including a “phone tree” that helps staff reach the right lawyer.
How Anthropic handled it: each team started from a human workflow it already ran:
- Growth Marketing already reviewed ad performance.
- Product Design already checked design states.
- Legal already routed incoming questions.
The agent shortened each workflow. It did not replace the workflow with something new. This matches the post’s thesis about focusing on the human workflows Claude Code can augment.
Do it this month:
- Ask each non-engineering lead for one annoying, repetitive workflow that involves files: CSVs, design files, contracts or docs. Workflows built around files give the agent something concrete to read.
- Pair the lead with an engineer for a 60-minute session to scaffold a first prototype. The engineer handles setup and permissions. The lead supplies domain knowledge.
- Keep the first prototypes internal and low-risk, as Legal’s routing tool was. A misrouted internal question costs little. A wrong answer sent to a customer costs a lot.
- Decide in advance who maintains the tool if it sticks. Prototypes that work become dependencies. Without an owner, they break quietly.
Claude Code for designers and Claude Code for legal teams can sound like a stretch. Anthropic’s examples show the real requirement is not coding skill. It is a clear, repeatable task and access to the right files.
How Can You Use Claude Code for Security Incident Response?
Give Claude Code the stack traces, relevant documentation and dashboard screenshots from the incident. Ask it to identify the likely cause and propose commands or fixes for a human to review.
How Anthropic handled it: Security Engineering fed in stack traces and documentation. It reports resolving problems 3x as quickly as the 10–15 minutes of manual scanning they typically took. The time savings came from skipping the reading step. The humans still owned the response.
Data Infrastructure followed the same pattern during a Kubernetes outage. The team shared dashboard screenshots, and Claude identified pod IP address exhaustion. It then supplied commands to create a new IP pool, which Anthropic says saved about 20 minutes. This was a single incident, not an average, but it shows how much a screenshot can carry when there is no time to write a structured log query.
Do it this month:
- Add an “AI triage” step to your incident runbook that lists exactly which artifacts to paste in: stack traces, the relevant service docs, dashboard screenshots and recent deploy notes.
- Run it on a past, resolved incident first. You already know the answer, so you can see whether the agent reaches it and how quickly.
- Require a human to review and run any infrastructure command the agent suggests. AI incident triage speeds up diagnosis. It does not replace change control.
- Record the agent’s diagnosis in the postmortem. Over a quarter, you will build an accuracy record that tells you how much weight to give it.
How Do You Use CLAUDE.md to Onboard New Engineers and Data Scientists?
Write CLAUDE.md files that explain your codebase’s structure and conventions. New hires can then point Claude Code at the codebase and those files and ask it to explain dependencies and unfamiliar functions.
CLAUDE.md is a Markdown memory file that Claude Code loads automatically to get project context and instructions. The /init command can generate a starting version. For current file locations, scoping rules and syntax, see the official Claude Code documentation. We won’t restate details here that may have changed.
How Anthropic handled it: new Infrastructure data scientists feed Claude Code the codebase plus CLAUDE.md files to understand data pipeline dependencies. On the Inference team, members without ML backgrounds use Claude Code to explain model-specific functions. Anthropic reports their research time falling from about an hour to 10–20 minutes.
Do it this month:
- Create a CLAUDE.md file in your most-onboarded repo. Cover the architecture, key directories, build and test commands, conventions and known gotchas. Keep it short. The agent needs signal, not a wiki dump.
- Give new hires a starter prompt list. For example: “Explain how X flows into Y.” “Which files would I change to fix Z?” “What breaks if I modify this function?”
- Have a senior engineer review the CLAUDE.md file each quarter. Treat it as living onboarding documentation with a named owner. Stale instructions are worse than none, because the agent will follow them with confidence.
What Are AI Coding Agent Use Cases by Team?
Match the use case to the team’s main bottleneck. For engineering, that is finding code. For security, it is triaging incidents. For infrastructure, it is onboarding. For marketing, it is analyzing and producing creative variations. For design, it is finding edge cases. For legal, it is routing requests.
The table separates what Anthropic reported from our suggestions. The right-hand column contains our recommendations, not Anthropic facts.
| Team | Main bottleneck | Proven at Anthropic | Adapt for your team (our recommendation) |
|---|---|---|---|
| Engineering | Finding the right code | “First stop” for locating files for bugs and features | Add a CLAUDE.md to every active repo. Make the agent step one in bug triage. |
| Security | Incident triage | Stack traces + docs → 3x faster resolution | Standardize which artifacts go into the agent during every incident. |
| Infrastructure / Data | Onboarding and outages | CLAUDE.md onboarding. Screenshot-based outage diagnosis. | Build a starter prompt pack for each new hire’s first two weeks. |
| Marketing | Creative volume and analysis | Two sub-agents over ad CSVs. Figma plugin with up to 100 variations per run. | Begin with one channel and one fixed performance threshold. |
| Design | Missed edge cases | Autonomous loops on Figma files. Error-state and logic-flow mapping. | Run an edge-case pass before every design review. |
| Legal | Routing requests | “Phone tree” routing prototype | Prototype intake forms that route to the right owner. Keep them internal. |
Results depend on how much context each team can give the agent. A team with clean CSVs and well-organized files will see faster gains than a team whose knowledge lives in people’s heads.
How Can Marketing Teams Use Claude Code?
Marketing teams can build agent workflows that ingest ad performance data, flag underperformers and generate new variations. Anthropic’s Growth Marketing team did exactly this.
How Anthropic handled it: Growth Marketing built an agentic workflow that processes CSV files containing hundreds of ads, flags underperformers and generates new variations. It uses two specialized Claude Code sub-agents. The post splits them by output type: one handles headlines and the other handles descriptions. Separately, a Figma plugin generates up to 100 ad variations per run. That is a capacity figure, not a performance result.
Why the split matters (our interpretation): headlines and descriptions have different character limits and different jobs. Giving each sub-agent one narrow task and its own context window reduces confusion between tasks and keeps outputs within format limits. The general lesson is that sub-agents work best with tight scopes and fixed output formats.
Do it this month:
- Export ad performance to a CSV with consistent columns. Use the same names, the same metrics and the same date ranges every time.
- Define “underperforming” numerically before the agent sees the data. For example: CTR below X over at least Y impressions. If you leave the threshold vague, the agent will invent one.
- Split the work into narrow jobs. One step flags ads against your threshold. Separate sub-agents draft headlines and descriptions within brand guidelines and character limits.
- Keep human review on all copy before it ships. Volume is the easy part. Brand voice and legal claims still need a person.
How Do You Roll Out Claude Code Beyond Engineering?
Start from existing human workflows, give each team good context, and let teams build small internal tools before you expand. Anthropic’s examples spread across teams this way rather than through a single top-down mandate. Teams as different as Legal and Growth Marketing found their own uses in workflows they already ran.
30-day rollout plan:
- Week 1: Pick 2–3 non-engineering pilot teams and one workflow each. Favor workflows that involve files and repeat weekly.
- Week 2: Prepare the context: CLAUDE.md files, sample data, design files, routing rules. This week sets the ceiling for everything after it.
- Week 3: Build and run prototypes with engineering support. Hold office hours instead of handing over a doc and leaving.
- Week 4: Measure results and decide which tools to keep. Document guardrails covering data access, review steps and ownership.
Set autonomy levels per use case. Autonomous loops suit low-risk, reversible iteration such as design checks. Incident commands need a human to approve them. This matters more as agentic coding workflows shift work from assistance to delegation. Anthropic’s Economic Index report on software development (April 2025) classified 79% of Claude Code conversations as “automation,” where the AI performs the task directly. The figure for Claude.ai was 49%. More delegation means review steps have to be designed on purpose, not added afterward.
What to Copy and What Not to Copy
Copy Anthropic’s context-first habit. Treat its numbers with care.
Copy
- Context first. Every Anthropic example begins with a codebase, CLAUDE.md files, stack traces, screenshots or source files. If you can’t name the artifact the agent will read, you aren’t ready to start.
- Workflow-first thinking. The post’s thesis is about augmenting human workflows. Rollouts that start with “find uses for AI” tend to stall. Rollouts that start with “what do we scan or search manually every week?” tend to stick.
Caveats
- Self-reported numbers. The 3x, 80%, 20-minute and up-to-100-variation figures all come from Anthropic about its own product. No independent benchmark exists.
- Unusual users. Anthropic’s staff know Claude better than almost anyone. Your results may differ.
- Dated tooling. The post is from July 2025. Features have changed since then, so check the current docs.
- Prototype status. The post describes the Legal phone tree as a prototype, not a production system.
Don’t copy
- Unreviewed infrastructure commands. Anthropic’s examples show diagnosis and suggested commands, not unsupervised production changes. Don’t read the Kubernetes story as permission to let an agent change live clusters.
Where Customer Evidence Fits: The Missing Context
Anthropic’s examples work because the agent gets good context: code, CLAUDE.md files, stack traces and screenshots. None of that context includes what customers actually asked for. This section is our view, separate from Anthropic’s story. Anthropic did not use or mention BuildBetter.
Context about the code is enough for debugging and onboarding. Once agents start building features, they also need requirements context: what customers said on calls, in support tickets and in Slack threads with account teams. Without it, an agent writes clean code that solves the wrong problem.
The BuildBetter MCP server gives coding agents that customer context. Claude Code, Cursor and other Model Context Protocol clients can query calls, signals, contacts and feedback directly, so the agent builds to real requirements instead of guesses.
BuildBetter is one source of truth for internal and external customer voice. It brings internal conversations (calls and Slack) together with external feedback (support tickets and surveys) through 100+ integrations, including Zoom, Jira, Zendesk and HubSpot. Each piece of feedback is analyzed individually for severity and business impact, using your own taxonomy.
A practical pattern:
- Pull the customer evidence for a feature through MCP: the quotes, the accounts and the severity.
- Have the agent draft a PRD or ticket grounded in that evidence, or generate one in BuildBetter Documents.
- Build against that spec, with the CLAUDE.md file supplying codebase context.
The flow runs from customer conversation to shipped decision, with evidence attached at every step.
An honest limit: BuildBetter is not built for distributing enterprise surveys at scale. If that is your main need, use a purpose-built survey platform.
FAQ
When did Anthropic publish how its teams use Claude Code?
On July 24, 2025, in a post titled “How Anthropic teams use Claude Code,” now hosted at claude.com/blog. The original anthropic.com/news URL redirects there.
How does Anthropic use Claude Code internally?
Across engineering and non-engineering teams:
- navigating codebases (Product Engineering’s “first stop”)
- triaging incidents (Security, Data Infrastructure)
- explaining unfamiliar code (Inference)
- onboarding new data scientists with CLAUDE.md
- generating ad variations (Growth Marketing)
- finding design edge cases (Product Design)
- prototyping routing tools (Legal)
Can non-engineers really use Claude Code?
Yes, according to Anthropic. Growth Marketing, Product Design and Legal all built workflows or prototypes. Each started from a workflow the team already ran and from files it already had, such as CSVs, Figma files and request types.
How much time did Claude Code save Anthropic’s teams?
Anthropic reports three results:
- about 3x faster resolution of incident problems in Security, compared with 10–15 minutes of manual scanning
- an 80% cut in research time in Inference, from about 1 hour to 10–20 minutes
- about 20 minutes saved in one Kubernetes outage
All figures are self-reported, with no independent benchmark.
What is CLAUDE.md and how is it used in onboarding?
CLAUDE.md is a Markdown file that Claude Code loads automatically to get project context and instructions. Anthropic’s new Infrastructure data scientists point Claude Code at the codebase plus CLAUDE.md files to understand data pipeline dependencies. See the official docs for current file locations and the /init command.
Should Claude Code run autonomously?
It depends on the risk. At Anthropic, design iteration ran in autonomous loops. In the incident examples, the agent suggested commands for humans to run.
How do I give a coding agent customer context?
Connect a customer-evidence source through MCP. BuildBetter’s MCP server gives agents access to calls, tickets and Slack feedback, so features are built against what customers actually said.
Sources and Further Reading
- Anthropic, “How Anthropic teams use Claude Code” (July 24, 2025): claude.com/blog/how-anthropic-teams-use-claude-code
- Claude Code official documentation: docs.anthropic.com/en/docs/claude-code/overview
- Model Context Protocol official site: modelcontextprotocol.io
- Kubernetes on Wikipedia (background for the outage example): en.wikipedia.org/wiki/Kubernetes
- Anthropic on Wikipedia: en.wikipedia.org/wiki/Anthropic
- BuildBetter: buildbetter.ai
Give Your Agents the Customer Context They’re Missing
Anthropic’s teams showed that coding agents do their best work when they get the right context. For product teams, that context has to include the customer. BuildBetter captures every call, ticket, Slack thread and survey and turns it into evidence your team and your agents can build from. That evidence produces PRDs, tickets and customer follow-ups, not dashboards no one opens.
