How the airline's engineers use Claude in the repo

What Claude Code does for the engineers

Claude Code is a coding agent that reads a codebase, edits files, runs commands and uses git.1 The same engine runs in the terminal, VS Code, JetBrains, the desktop app, the web and the phone, and repository settings carry across the local ones.2 At the fictional airline it builds the rebooking assistant from the signed spec: the refund tool with the $400 limit in code, its tests, and the evaluation on the 500 past cases. Arjun, the architect, owns the instructions it reads. Maya, the QA lead, owns the tests that judge what it writes.

When to reach for something else

• Before the spec is signed. A coding agent builds what the document says, gaps included, and the lab Grow the spec shows what one guesses from a paragraph.
• When the airline reaches Claude only through its AWS account or an API key. With an API key or a cloud provider such as Amazon Bedrock you lose cloud sessions, routines, Remote Control and the Chrome integration, and on Bedrock the built-in web search too.3
• When the job is a document, a workbook or a chase. That is Claude at the desk, the first manual.

Five moves the airline's engineers make with it

Each move names who makes it, in which phase, and the words they type. Copy one and change the nouns.

1. Move 1P1Arjun, architect

   Write the file every session reads

   CLAUDE.md is a plain file of instructions loaded at the start of every session, at organisation, user, project or local scope.4 Arjun writes the project's file before the first session, and reviews changes to it like code.

   If a repository has AGENTS.md and no CLAUDE.md or CLAUDE.local.md, Claude Code reads AGENTS.md.5

   CLAUDE.md Copy

   # Rebooking service
   
   The spec is docs/spec.md. Read fields 3, 5 and 7 before you change behaviour.
   The $400 refund limit is one constant in refund_tool.py. Never copy it into a prompt.
   Every limit in the spec's field 5 is enforced in code, with a test below, at and above it.
   Never book on a stale fare: when the fare engine is down, the case queues for an agent.
   Run the tests before you call a change done.

2. Move 2P2An engineer

   Ask for the plan before the code

   In plan mode Claude Code researches and proposes, and edits stay blocked until you approve the plan, except in a terminal session where bypass permissions are available.6 Before the refund tool is touched, the engineer asks what Claude Code would build, and what it would have to guess.

   Every guess on the third list goes back to its owner before an edit is approved.

   The lab's last step runs this prompt on a paragraph and on the spec, with recorded replies →

   The hand-off prompt, from the lab Copy

   You are a coding agent about to implement the feature described below. Before you write any code, answer in three short lists:
   1. The tools or functions you will build.
   2. Every limit you will enforce in code, each with the exact line of the document it comes from.
   3. Everything you would have to guess, because the document does not say.
   
   DOCUMENT:
   <the signed spec: fields 3, 7 and 8>

3. Move 3P2Maya, QA lead

   Review the branch against the file

   The command /code-review runs a review in a background subagent of your session and follows CLAUDE.md.7 Maya runs it, then asks the question a review rarely asks.

   At low and medium effort /code-review reports only the findings it is most confident in, and at high and above it may include findings it is less sure about.8 So the review is a second reader, and the test at $400.01 is the gate.

   In the session Copy

   /code-review
   
   List every number in this branch that also appears in docs/spec.md.
   For each, say whether it is read from one constant or typed again.
   The $400 refund limit must appear once, in refund_tool.py.

4. Move 4P2An engineer

   Hand the evaluation run to a cloud session

   A cloud session runs on an isolated machine and keeps going after you close the laptop: start one with claude --cloud "task" and pull it back with claude --teleport.9 The run on the 500 past cases takes a while, so the engineer sends it away.

   The push comes first. The cloud machine clones your GitHub remote, so commits you have not pushed are not there.10

   Terminal Copy

   git push
   claude --cloud "Run the evaluation on the 500 past cases in eval/cases.csv. Report the share handled right for each kind of case, codeshare as its own group, against the bar in docs/spec.md field 6. Open a pull request with the report. Change no code."

5. Move 5P2Maya, QA lead

   Walk the agent's screen in Chrome

   Claude in Chrome is a browser extension that reads pages, clicks, types, fills forms and moves between tabs, and its Chrome side panel is in beta.11 Maya checks what a contact-centre agent sees when the refund path runs.

   It acts directly on websites on your behalf, and the help page says plainly that, despite its safety classifiers, it is still risky.12 So Maya runs it in a browser profile signed in to the test site and nothing else.

   Terminal, then the prompt Copy

   claude --chrome
   
   Open the rebooking screen on the local build as the test agent.
   Cancel booking TEST-104 and take the refund path with a $399 fare, then with a $401 fare.
   Report what the screen shows at each step, and every console error.

Where Claude Code lets you down

• Its undo is partial. The command /rewind restores code, conversation or both from a snapshot taken at each prompt, but changes made by shell commands are not tracked, and it is not version control.13 Commit before every long run.
• Auto mode is a convenience. The docs say auto mode cuts prompts but does not guarantee safety, and an owner can switch it off with disableAutoMode.14
• A helper does not hear the conversation. A subagent has its own context window, system prompt and tools, is defined as a markdown file in .claude/agents/, and returns a summary.15 A subagent other than a fork does not see your conversation or the files Claude has read, and starts from the brief Claude writes, its own prompt and your CLAUDE.md files.16 Write the brief as you would for a contractor on their first day.
• Its sandbox has edges. The sandbox limits which files and hosts shell commands can reach, and it does not cover file tools, MCP servers or hooks.17

Three settings decide how it behaves

The traps that catch a team

• A one-line job in CI can run more than you meant. The command claude -p "prompt" runs once and exits, and --bare skips hooks, skills, plugins, MCP servers, memory and CLAUDE.md so every machine behaves the same.25 Without --bare, a -p run executes the repository's hooks and connects its MCP servers with no trust prompt.26
• Routines run a saved task in the cloud on a schedule, from an HTTP call or from a GitHub event, and act as you on GitHub and in your connectors.27 A green run status means the session ran, not that the task succeeded.28 Read the pull request, not the tick.
• Auto-fix can answer review comments under your GitHub name, which can set off automation that is driven by comments, such as Terraform runs.29
• The command claude --worktree name gives a session its own checkout and branch, and ignored files such as .env are missing unless listed in .worktreeinclude.30
• Steering from the phone still needs your own machine. Remote Control lets you drive a session running on your own machine from claude.ai/code or the phone, as long as the local process stays alive.31

Where every fact on this page comes from

Each numbered mark above points to a row here: the fact, the vendor's page it was checked against, and the date. A date turns amber once it is more than 60 days old. Prices, usage limits and model names change too often to print: read them on Claude's plans and prices and the models overview.