01How Anthropic Engineers Use Claude Code for Real Work
Learning how anthropic engineers use claude code shows you that it is a business tool as much as a developer tool. Claude Code is the coding agent Anthropic's own engineers use every day, and Julian Goldie uses it to build lead trackers, client reports, booking pages and internal dashboards.
Anthropic published how its own teams use Claude Code. Its growth marketing team, for example, feeds in spreadsheets of ad results and lets two helper agents write new versions. Its product engineers ask which files matter before they start.
In this episode, Julian takes nine Claude Code features and turns each one into a real business workflow with the exact prompt he types:
- A lead watcher and a morning client report.
- A finish line that Claude works toward on its own.
- A quality guard that checks every report.
- Three versions of a page built at once.
- A site wide update in one prompt.
- Control from your phone, code review and a team of specialist agents.
02Workflow 1: The Lead Watcher With /loop
Every business has a place where new enquiries land, like a form, an inbox export or a folder of files. Somebody has to check it. The /loop command makes Claude do it on a timer.
If your leads live in a tool like your CRM, connect it first with claude mcp add. That gives Claude a safe connection to that tool.
- Connect your lead tool with claude mcp add.
- Type /mcp to check it is connected and signed in.
- Type /loop, then how often, then the job. Julian's prompt: /loop five minutes, check the leads folder for new enquiries, score each one hot, warm or cold, and add one line per lead to leads.md.
- Claude confirms the timing and gives the job an ID.
Every five minutes, it wakes up, reads the new files and the list grows on its own.
One thing to know: close the laptop and it stops. The loop runs on your own computer, so for jobs that must run while you sleep, you need the next workflow.
03Workflow 2: The Morning Client Report With /schedule
This one runs in the cloud, so your computer can be switched off. You type /schedule and describe the job in plain words.
Julian's prompt: /schedule every weekday at 8am, read yesterday's traffic export, compare it to last week and write a one page client report with three wins and one thing to fix.
Claude does not just save it. It asks you questions first:
- When should it run?
- Which project should it use?
- What exactly should the prompt say?
Then it saves it as a scheduled task. A scheduled task can start on a timer, from an API call or from something happening on GitHub. Anthropic calls this a research preview right now, so expect it to keep changing.
If you already run automation somewhere else, there is print mode. You type claude -p, then your prompt in quotes. It runs once and prints the answer.
Add the output format JSON option and your other tools can read the result.
So with two workflows, one watches your leads while you work and one writes your client report before you wake up.
04Workflow 3 and 4: /goal, Rewind and Hooks
With /goal, you set a finish line, such as an importer that loads the sample contacts file with zero errors and passes every test. You see goal active at the bottom. Claude keeps taking turns toward that condition, and a separate evaluator checks if it is met.
It only stops when the check passes.
- Type /goal to see the status.
- Type /goal clear to remove it.
- Press escape twice to open rewind and restore code, the conversation or both.
One honest note: rewind tracks the file edits Claude made. It will not undo something already sent to an outside system, so test on sample data first.
A hook is a command that runs automatically at a set moment: before a tool runs, after a tool succeeds or when Claude stops. Julian adds a post tool use hook in the settings file. The matcher says write or edit, and the command runs his report checker.
If a report is missing the client's name or the date range, it gets caught the moment the file is saved.
He also adds a stop hook that runs the tests, so he never gets a done message on broken code. Type /hooks to see every hook you set up.
05Workflow 5 and 6: Worktrees and /batch for Big Jobs
Say a client wants three versions of a landing page. Normally you build one, then the next, then the next. With worktrees, you run three sessions side by side and none of them touch each other.
- Make sure the project has at least one git commit, or worktrees will not work.
- Start each session with claude -w and a name, such as landing A, landing B and landing C.
- Each copy gets its own branch and lives in its own folder inside .claude/worktrees.
- Compare all three, keep the winner and merge it.
/batch is Julian's favorite. Say a client has a booking form on every location page and it needs replacing everywhere. You type /batch and ask Claude to replace the old booking form with the new one on every location page in the pages folder.
Claude splits the job into independent pieces, somewhere between five and 30. It shows you the plan first, and you approve it. Each piece then gets its own background agent in its own worktree, and the results come back one by one.
The same trick works for new footer text on every page, a new field on every form or a renamed product. If you would copy and paste it 30 times, that is a batch job.
06Workflow 7 and 8: Remote Control and Code Review
Long builds do not need you at the desk. In your project, type claude remote-control, or inside a session type /remote-control. Claude shows a connection link.
Open it in the Claude app on your phone, and the conversation appears there live.
The work still runs on your own computer, so leave it switched on. This pairs well with /goal. Set the finish line, start remote control and walk away.
When Claude needs a decision, you answer from your phone and the build keeps moving.
Before anything ships, review it. Before Julian's lead routing changes go live, he runs these checks:
- /code-review: a review agent reads every change on the branch in the background, and findings come back into the chat. Add the fix option and it fixes them too.
- /security-review: checks the same changes for risks like injection, broken logins or customer data being exposed.
- Code review on pull requests: for teams on GitHub, you comment to ask Claude for a review, and agents post findings on the lines of code, ranked by severity.
07Workflow 9: Specialist Subagents Like Anthropic's Team
This is the closest copy of what Anthropic's growth marketing team does. One agent writes, and another agent checks.
- Create a file in .claude/agents called reportchecker.md.
- At the top, add a name and a short description.
- Underneath, write its instructions: check every number against the source export and flag anything without a date range.
- Ask Claude to use the reportchecker agent to review your report.
You will see a row show the agent's name and its task. The subagent works in its own context, so your main chat stays clean. The checker comes back with a short list of problems for you to fix.
Why this helps:
- The writer and the checker do not mix up their jobs.
- Your main conversation does not fill up with checking notes.
- You can reuse the same checker on every report.
You can use the same pattern for any job where one pass writes and a second pass checks, such as client reports built from a traffic export.
08Make Every Claude Code Workflow Repeatable With Skills
The part that makes all nine stick is skills. Every workflow you like, you save as a skill. A skill is a folder with a SKILL.md file inside that holds the steps, the format and the checks.
Next time, you type /client-report and Claude follows the exact same process. For the rules about the client, like the tone, brand colors and report format, type /init to create your project memory and /memory to see what Claude remembers.
Here is the whole stack in one list:
- /loop watches and /schedule reports.
- /goal finishes the job and hooks guard quality.
- Worktrees split versions and /batch scales changes.
- Subagents split the work and remote control lets you step away.
- Code review protects the client and skills make it repeatable.
This is for business owners and teams who want reports, lead tracking and site updates done for them, not just developers. Start with one workflow, test it on sample data, and save it as a skill once it works.
