SkyWaysThe agentic manual

Tool guide 1 · the Claude apps

How the airline's team uses Claude at the desk

Chat, Projects, skills, connectors and scheduled tasks, for the people who write the spec.

5 moves3 settings8 traps31 dated factsLast checked

On this page
  1. What it is
  2. When not to use it
  3. Five moves
  4. Where it falls short
  5. Three settings
  6. Its traps
  7. The facts, dated

What the Claude apps do for a team

Claude at the desk means the Claude apps, on the web, the desktop and the phone. Cowork, the separate agent mode for multi-step knowledge work, is being folded into ordinary chat: you describe the outcome and Claude picks the tools.1 Since 16 September 2026 the agent features of Cowork have been rolling into ordinary chats, starting with the Pro and Max plans.2 Around the chat sit the four things this manual covers: projects, skills, connectors and scheduled tasks. At the fictional airline, this is where Priya, the product manager, writes the rebooking assistant's spec, and where Maya, the QA lead, reads the 500 past cases.

When to reach for something else

  • When the work is code in a repository. That is Claude in the repo, the second manual.
  • When a rule must hold every time. Project instructions tailor Claude's replies within the project, for example a more formal tone or the view of a particular role.3 The $400 refund limit belongs in the refund tool's code, never in a project's instructions.
  • When you are tuning the rebooking assistant's own system prompt. The Console's Workbench is now the playground, which keeps your current draft in your browser and has no saved prompts or prompt versions.4 Keep that prompt in git, beside the tests that judge it.
  • When the system you need sits inside the airline's network. A custom connector is called from Anthropic's cloud, so its server must be reachable from the public internet, and one behind a VPN will not connect.5

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 1P1Priya, product manager

    Keep the whole case in one project

    Priya opens a project for the rebooking assistant and loads the PRD-lite, the interview notes and the glossary. A project is a self-contained workspace with its own chats and knowledge base, where you upload documents and set the project's own instructions.6 She writes its instructions once, so no chat starts cold.

    Then she shares it with Arjun, the architect, and Maya. On Team and Enterprise a project can be shared with view or edit permission, and the members it is shared with see its knowledge and instructions.7

    Project instructions
    This project is SkyWays' rebooking assistant. Owner: Priya (product).
    Read the PRD in the project knowledge before you answer.
    Take every number from the PRD: the 38-minute wait, the 240 cases a day, the $400 refund limit.
    Never supply a number yourself.
    Where the PRD is silent, write NOT DECIDED, the question, and who answers it.
  2. Move 2P1Priya, product manager

    Ask for the questions before the draft

    A draft fills every gap with a guess that reads like a decision. So Priya first asks what her one page leaves open, and who owns each answer.

    The lab Grow the spec opens on this prompt, with a real model's recorded reply

    Prompt A, from the lab
    Below is a one-page PRD-lite for a rebooking assistant at an airline. Do not write the PRD yet. List the questions this page does not answer and a build could not start without. Most blocking first, eight at most, one line each. After each question, name the role that usually owns the answer.
  3. Move 3P1Arjun, architect

    Write the spec's rules down once, as a skill

    The eight-field spec is a procedure, and a procedure retyped into every chat drifts. A skill is a folder of instructions, scripts and resources that Claude loads only when it judges the skill relevant to the task.8 Arjun writes one for the whole team.

    Skills need code execution switched on, and a skill can run code and install packages.9 Arjun reads a skill from outside the team before he switches it on, as he would any software.

    The lab's second prompt teaches the same rule, and shows what a model writes without it

    SKILL.md
    ---
    name: eight-field-spec
    description: Turn a PRD into SkyWays' eight-field agent spec. Use when someone asks for a spec, or for the fields a spec still lacks.
    ---
    Write exactly eight fields: Title, Value, Acceptance criteria, The model's role, Autonomy, The bar, Fallback, Records.
    For fields 4 to 8, where the PRD does not say, write NOT DECIDED, the question, and who should answer it.
    Never fill a gap with a sensible default. The gaps are what the review is for.
  4. Move 4P2Maya, QA lead

    Turn the 500 past cases into a workbook

    Maya has the 500 past cases as an export: the kind of case, the old outcome, and whether it was right. Claude writes and runs code in a private sandbox to make real files: Excel, PowerPoint, Word, PDF and charts.10

    She checks three totals by hand before any number goes near the bar. The help page itself warns that a hidden instruction in a file or page could make Claude run untrusted code or send data out.11 So she works from the airline's own export, never from a file a stranger sent.

    Prompt, with the CSV attached
    The attached CSV holds 500 past rebooking cases, one row each.
    Make an Excel workbook with one sheet per kind of case: same-day SkyWays, partner (codeshare), refund, hand-over to an agent.
    On each sheet show the count, the share handled right, and the formula behind every number.
    Keep codeshare as its own group.
    Do not drop or repair any row. List the rows you could not read.
  5. Move 5P1Priya, product manager

    Chase the open questions every morning

    A spec waits on its owners, and chasing them is a daily chore. A scheduled task is a saved prompt that Claude runs hourly, daily, weekly, on weekdays or on demand, in the cloud, so your computer can be off.12 A connector lets Claude reach another service, either one from the directory, such as Slack, Google Drive or Linear, or a custom one that points at a remote MCP server.13

    When you set up a scheduled task by hand, you choose its approval mode, along with its prompt and how often it runs.14 Priya's task may read, and may not post.

    Scheduled task, weekdays at 08:00
    Read the rebooking assistant's board in Jira and the #rebooking channel in Slack.
    List every spec field still marked NOT DECIDED, its owner, and the days it has been open, oldest first.
    Do not answer any of them. Do not post anywhere. Send the list to me.

Where the apps let you down

  • It advises and does not enforce. A project's instructions tailor what Claude writes3, so a limit that matters lives in code.
  • Its memory is not a record. Claude saves short topics about you and your work as you chat, and you can read, edit and delete each one in Settings.15 Put a decision in the project's files, with a name and a date beside it.
  • A citation shows where to look and proves nothing on its own. Research runs many searches that build on each other, across the web and your connected sources, and returns a cited report in minutes.16 Open the sources that carry a decision before you quote the report.

Three settings decide how it behaves

1

Memory: what it keeps between chats

Memory is on by default on individual plans; on Team and Enterprise it stays off until an owner allows it and each member turns it on.17 Incognito chats stay out of history and memory, but on Team and Enterprise they are still kept for a period and appear in an owner's data exports.18

2

Permissions: what it may touch

On Team and Enterprise an owner enables a connector, each member signs in, and the owner can limit read, write and delete actions for everyone.19 A connector inherits each person's own permissions in the service it reaches, except a custom connector on a shared credential, which reaches whatever that credential can.20 On Team and Enterprise an owner chooses what the sandbox may reach, and the help page says network access starts off, though for Team it also says it starts at package managers only.21

3

Context: what every chat starts from

On paid plans, when a project's knowledge nears the context limit, Claude switches to retrieval and the project can hold far more.22 Load the PRD and the decisions, and leave the email threads out.

The traps that catch a team

  • If an owner switches memory off for the organisation, every member's memories are deleted at once.23 Keep decisions in files, where that switch cannot reach them.
  • In the merged chat the GitHub integration and conversation branching are missing for now, and incognito chats fall back to the old experience.24
  • A long task runs in Anthropic's cloud, labelled beta, and reaches your computer through the desktop app only when it needs local files, the browser or computer use.25
  • The help centre warns that custom connectors reach servers Anthropic has not verified, and advises connecting only to servers from trusted organisations.26
  • Two of Anthropic's own pages disagree on who can share a skill. Anthropic's platform docs say custom skills in the apps are per user and cannot be managed centrally.27 The help centre says owners on Team and Enterprise can give a skill to everyone in the organisation.28
  • A scheduled task works with your connectors and the files in your Claude account, and one that needs local files or apps runs only locally.29
  • Even in a shared project, each member's chats stay private unless that member shares one, which creates a link to it.30 So a decision reached in one member's chat belongs in the project's files, where the others can read it.
  • For now, two things share one name. Claude Code's projects, a public beta, run related tasks as parallel cloud sessions, while projects in chat and Cowork are the earlier experience and keep working until the redesign reaches them.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.

Every fact this page uses, with its source and the date it was checked
No.The factSource, and when it was checked
1Cowork, the separate agent mode for multi-step knowledge work, is being folded into ordinary chat: you describe the outcome and Claude picks the tools.support.claude.com/16761823-claude-cowork-and-chat-are-one-claude
2Since 16 September 2026 the agent features of Cowork have been rolling into ordinary chats, starting with the Pro and Max plans.support.claude.com/12138966-release-notes
3Project instructions tailor Claude's replies within the project, for example a more formal tone or the view of a particular role.support.claude.com/9517075-what-are-projects
4The Console's Workbench is now the playground, which keeps your current draft in your browser and has no saved prompts or prompt versions.support.claude.com/8606378-how-do-i-use-the-workbench
5A custom connector is called from Anthropic's cloud, so its server must be reachable from the public internet, and one behind a VPN will not connect.support.claude.com/11175166-get-started-with-custom-connectors-using-remote-mcp
6A project is a self-contained workspace with its own chats and knowledge base, where you upload documents and set the project's own instructions.support.claude.com/9517075-what-are-projects
7On Team and Enterprise a project can be shared with view or edit permission, and the members it is shared with see its knowledge and instructions.support.claude.com/9519189-manage-project-visibility-and-sharing
8A skill is a folder of instructions, scripts and resources that Claude loads only when it judges the skill relevant to the task.support.claude.com/12512176-what-are-skills
9Skills need code execution switched on, and a skill can run code and install packages.support.claude.com/12512180-use-skills-in-claude
10Claude writes and runs code in a private sandbox to make real files: Excel, PowerPoint, Word, PDF and charts.support.claude.com/12111783-create-and-edit-files-with-claude
11The help page itself warns that a hidden instruction in a file or page could make Claude run untrusted code or send data out.support.claude.com/12111783-create-and-edit-files-with-claude
12A scheduled task is a saved prompt that Claude runs hourly, daily, weekly, on weekdays or on demand, in the cloud, so your computer can be off.support.claude.com/13854387-schedule-recurring-tasks-in-claude-cowork
13A connector lets Claude reach another service, either one from the directory, such as Slack, Google Drive or Linear, or a custom one that points at a remote MCP server.support.claude.com/11176164-use-connectors-to-extend-claude-s-capabilities
14When you set up a scheduled task by hand, you choose its approval mode, along with its prompt and how often it runs.support.claude.com/13854387-schedule-recurring-tasks-in-claude-cowork
15Claude saves short topics about you and your work as you chat, and you can read, edit and delete each one in Settings.support.claude.com/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context
16Research runs many searches that build on each other, across the web and your connected sources, and returns a cited report in minutes.support.claude.com/11088861-use-research-on-claude
17Memory is on by default on individual plans; on Team and Enterprise it stays off until an owner allows it and each member turns it on.support.claude.com/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context
18Incognito chats stay out of history and memory, but on Team and Enterprise they are still kept for a period and appear in an owner's data exports.support.claude.com/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context
19On Team and Enterprise an owner enables a connector, each member signs in, and the owner can limit read, write and delete actions for everyone.support.claude.com/11176164-use-connectors-to-extend-claude-s-capabilities
20A connector inherits each person's own permissions in the service it reaches, except a custom connector on a shared credential, which reaches whatever that credential can.support.claude.com/11176164-use-connectors-to-extend-claude-s-capabilities
21On Team and Enterprise an owner chooses what the sandbox may reach, and the help page says network access starts off, though for Team it also says it starts at package managers only.support.claude.com/12111783-create-and-edit-files-with-claude
22On paid plans, when a project's knowledge nears the context limit, Claude switches to retrieval and the project can hold far more.support.claude.com/11473015-retrieval-augmented-generation-rag-for-projects
23If an owner switches memory off for the organisation, every member's memories are deleted at once.support.claude.com/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context
24In the merged chat the GitHub integration and conversation branching are missing for now, and incognito chats fall back to the old experience.support.claude.com/16761823-claude-cowork-and-chat-are-one-claude
25A long task runs in Anthropic's cloud, labelled beta, and reaches your computer through the desktop app only when it needs local files, the browser or computer use. betasupport.claude.com/13345190-get-started-with-claude-cowork
26The help centre warns that custom connectors reach servers Anthropic has not verified, and advises connecting only to servers from trusted organisations.support.claude.com/11176164-use-connectors-to-extend-claude-s-capabilities
27Anthropic's platform docs say custom skills in the apps are per user and cannot be managed centrally.platform.claude.com/overview
28The help centre says owners on Team and Enterprise can give a skill to everyone in the organisation.support.claude.com/12512180-use-skills-in-claude
29A scheduled task works with your connectors and the files in your Claude account, and one that needs local files or apps runs only locally.support.claude.com/13854387-schedule-recurring-tasks-in-claude-cowork
30Even in a shared project, each member's chats stay private unless that member shares one, which creates a link to it.support.claude.com/9519189-manage-project-visibility-and-sharing
31Claude Code's projects, a public beta, run related tasks as parallel cloud sessions, while projects in chat and Cowork are the earlier experience and keep working until the redesign reaches them. public betacode.claude.com/claude-projects

The same team, in the code: the second manual.