How the airline's team uses Google AI Studio and Jules

What AI Studio and Jules do for a team

AI Studio's playground takes a system instruction and a chat with a chosen model, with run settings at the side and a Get code button.1 Jules is an experimental coding agent that works on a GitHub repository in a virtual machine while you move on, and shows a plan for you to approve before it changes code.2 At the fictional airline, an engineer tries the rebooking assistant's instructions in AI Studio before anyone writes the call, and the team hands Jules the chores the signed spec makes plain, such as a missing test.

When to reach for something else

• When a prompt would carry a real passenger's details. Under the Gemini API terms, Google uses what you send to unpaid services such as AI Studio to improve its products, and human reviewers may read it.3 The terms say not to submit sensitive, confidential or personal information to unpaid services.4 AI Studio falls under the paid terms when the account can reach a Cloud project with active billing, or is a Workspace enterprise account.5 Until the account is on the paid terms, the team tests with made-up cases.
• When the code is not on GitHub. Jules reaches repositories only through GitHub for now, and its docs say other version control systems will follow.6
• When the check needs the service running, as the walk through the agent's screen does. Jules's setup scripts do not currently support long-lived processes such as a dev server or a watch script.7 That walk belongs to a browser agent, as in Claude in the repo8.
• When the team needs a paid plan on its company accounts. Paid Jules plans come through a Google AI plan, which for now is open only to personal Google accounts ending in @gmail.com.9

Five moves the airline's team makes with it

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

1. Move 1P1An engineer

   Try the assistant's instructions before any code

   The engineer opens the playground, pastes the draft into System Instructions, and plays a passenger whose flight is cancelled. AI Studio's Run settings panel holds the model's parameters and safety settings, and switches for tools such as structured output, function calling, code execution and grounding.10

   In AI Studio a system instruction can be changed after the chat has started.11 So the engineer starts a new chat for each version, and keeps the version that passed in git.

   System instructions Copy

   You are SkyWays' rebooking assistant. The passenger's flight is cancelled.
   Offer the next two SkyWays flights that have seats. Never invent a flight or a fare.
   For a refund, call the refund tool and offer only what it returns. Never name an amount yourself.
   For a partner (codeshare) booking, hand the case to an agent.
   When the fare engine is down, say so, and queue the case for an agent.

2. Move 2P1Arjun, architect

   Pin the reply to a schema

   The refund tool reads a record, not prose. Structured output makes the reply follow a JSON Schema you supply, and it supports a subset of JSON Schema.12 Arjun turns structured output on in Run settings and gives it the record the tool expects.

   Google's guide says structured output is syntactically correct JSON, and that you should always validate its values in your application.13 So the refund tool still checks every amount, and the schema fixes only the shape.

   Structured output schema Copy

   {
     "type": "object",
     "properties": {
       "action": {"type": "string", "enum": ["rebook", "refund", "hand_to_agent", "queue"]},
       "flight": {"type": ["string", "null"], "description": "A SkyWays flight number from the tool, or null"},
       "reason": {"type": "string", "description": "One sentence a contact-centre agent can read"}
     },
     "required": ["action", "flight", "reason"]
   }

3. Move 3P1Priya, product manager

   Ask with Search on, and keep the sources

   Before the fallback field is signed, Priya asks what the rules in SkyWays' markets require, with each claim tied to its source. With the Gemini API's Google Search tool the model runs the searches itself and answers with citations that link each span of text to its source.14

   With Grounding with Google Search, Google stores your prompts, the context you give and the output for thirty days to build the grounded results.15 So the prompt names markets, never a passenger.

   Prompt, with Grounding with Google Search on Copy

   SkyWays flies in these markets: <the list in section 2 of the PRD>.
   For each market, what must an airline offer a passenger whose flight it cancels: rebooking, a refund, care?
   Use the regulator's own pages and the law itself, nothing else.
   Cite every claim. Where a rule is unclear, or two sources disagree, say so.
   Do not estimate any amount.

4. Move 4P2An engineer

   Hand Jules the missing tests, and read the plan

   The spec wants the refund tool tested at and around its limit, and two of those tests are missing: a bounded chore. Jules Tools, installed with npm install -g @google/jules, starts a session with jules remote new and fetches its changes with jules remote pull.16

   Jules writes a plan before any code, and you can change it through the chat before you approve it.17 The engineer approves it only when the plan reads the limit from the constant.

   Terminal Copy

   jules remote new --repo skyways/rebooking \
     --session 'Add tests for refund_tool.py at the limit, one cent below it and one cent above it. Read the limit from REFUND_LIMIT and never type the number. Change nothing outside tests/. If a test fails, stop and say why.'

5. Move 5P2Maya, QA lead

   Start Jules from the bug report

   When a past case fails the new evaluation, Maya files it as an issue. Adding the label jules to a GitHub issue starts a task, and when Jules finishes it posts a link to the pull request on the issue.18

   When Jules publishes a branch it is the commit author, and when it publishes a pull request it appears as its creator.19 So Maya reviews the pull request as she would a stranger's.

   GitHub issue, labelled jules Copy

   Title: Codeshare refunds above the limit skip the agent
   
   Past case 318: a partner (codeshare) booking, and the passenger asks for a refund above the limit.
   Expected: the case goes to an agent, as the spec says.
   Actual: the assistant offers the refund itself.
   Change only the codeshare branch in rebooking/route.py.
   Add case 318 to eval/cases.csv. Do not touch refund_tool.py.

Where AI Studio and Jules let you down

• A long chat fills up. Every message in an AI Studio chat is included in the prompt, so a long conversation can reach the model's token limit.20
• An unread plan still runs. If you navigate away, Jules eventually approves its own plan on a timer.21 So a task is started only when someone can read its plan.
• Its machine is online. Jules runs code in a cloud machine with internet access, and its FAQ says not to commit secrets such as API keys to the repository.22

Three settings decide how it behaves

The traps that catch a team

• Two of Jules's own pages disagree on how finished it is: the getting-started page calls it experimental2. The Jules FAQ says Jules is currently in public beta.29
• Jules now fixes failing CI checks on the pull requests it creates and resubmits them, with no one asking.30 So a green check on a Jules pull request may be its second attempt. Read the commits, not the tick.
• Editing a scheduled task in Jules is not currently supported, so a change means deleting it and creating a new one.31
• A low temperature is not the safe setting it looks like. Google's prompting guide strongly recommends leaving temperature, top P and top K at their defaults for the current generation of Gemini models, warning that a temperature below 1.0 can cause looping.32

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 Jules's plans and limits, AI Studio's plans and the Gemini models page.