Operate Any Computer App From Your Phone With Dispatch

Operate any computer app from your phone with Dispatch — skill template

Turn this template into a working skill by resolving every TODO. The seed prompt below is the published Anthropic example this template was derived from; your skill should generalize it for repeated, hands-off use.

Outcome

In Claude Cowork, Dispatch with computer use lets Claude control your computer's mouse and keyboard from the Claude mobile app to work in apps that have no other interface Claude could reach.

TODO: Restate the outcome for YOUR context in one sentence: who runs this, on what input, producing what artifact, how often.

Seed prompt (verbatim from source, © Anthropic PBC)

Open my accounting app and pull the outstanding invoices report for March. List every invoice over 60 days past due with the client name, invoice number, and amount. Send me the totals by client.

Required context and inputs

• Computer use toggle enabled in Dispatch settings
• Desktop accounting application open or pinned to taskbar/Dock
• Optional: Gmail or Slack connector for routing results
• TODO: exact file paths / folders / message formats this skill should expect
• TODO: domain context the model cannot infer (naming conventions, thresholds, house style)

Tools, connectors, APIs, and authentication

• Gmail
• Slack
• TODO: confirm which connectors are enabled in the runtime that will execute this skill
• Authentication: connectors authenticate via their own OAuth flows — this skill must

NEVER ask for, store, or echo credentials, tokens, or API keys. If auth is missing, stop and tell the user which connector to enable.

Permissions and sensitive actions

• Reads: TODO (folders, channels, records this skill may read)
• Writes: TODO (what it may create/modify, and where)
• Held back for the primary agent / human: sending external communications, financial

transactions, deleting or overwriting originals, submitting web forms

Workflow

1. Describe the task — enable computer use in Dispatch settings; Claude asks for approval before controlling applications
2. Give Claude context — confirm the desktop application is open and computer use is enabled; no connectors required
3. What Claude creates — Claude opens the app, navigates to reports, runs the report, and delivers results
4. Follow up prompts — check web portals in Chrome, review reports at your desk, or email results to your team

TODO: adapt the steps above (from the source page) into imperative instructions for the executing agent, including what to do between steps.

Decision points

• TODO: list each point where the skill must choose between paths, with the rule to apply
• Default rule: prefer the reversible option; when two readings of the input are

plausible, surface both rather than picking silently.

Validation criteria

• Output matches the outcome statement above (spot-check against the seed prompt's asks)
• Every factual claim is traceable to a provided input, connector record, or cited source
• TODO: add one domain-specific check a reviewer in your org would apply

Failure modes and fallbacks

• Connector unavailable or unauthenticated → pause, tell the user exactly which connector to enable and why; offer a degraded run from uploaded files
• Expected files missing from the connected folder → list what was found, ask before proceeding on partial inputs
• Ambiguous or missing input → ask one targeted question; if unattended, state the assumption inline and proceed
• Any step would take a sensitive/irreversible action → stop and hand back to the user (see Delegation)

Delegation

Apply the repo's delegation policy (docs/delegation-policy.md — bundle or restate it if you install this skill outside the repo). Defaults for this template:

• Run single-agent unless a step fans out over independent items (files, records,

vendors, channels). Only independent work parallelizes.

• Each delegated task must ship with: the minimal context slice it needs, an explicit

output contract, a validation check the primary agent runs on the result, and a fallback if it returns empty or fails.

• Final review, synthesis, and every sensitive action listed above stay with the

primary agent.

• TODO: name the concrete subtasks (if any) that qualify for delegation here.

Attribution

Derived from Operate any computer app from your phone with Dispatch (retrieved 2026-07-26). Seed prompt and workflow content © Anthropic PBC. Scaffold structure original to this repository.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: iankiku
• Source: iankiku/agent-skill-factory
• License: MIT

Install and usage instructions live in the source repository linked above.