Advanced
📅 2026-10-11 ⏱️ 9 min read Dean Dean

Automate Multi-Step Tasks on Android With Confirmation and Recovery

Choose a FoneClaw Workflow or Shortcut, review a meeting Do Not Disturb routine, update saved actions, and verify results under actual permissions and approval settings.

More videos on this topic

Videos reflect their publication dates. See this article for current steps and conditions.

  1. FoneClaw: Your Android AI Agent | Less Tapping, More Doing

    0:32Shorts

  2. AI Assistant for Android That Changes Phone Settings | FoneClaw

    0:47

  3. Change Phone Settings with FoneClaw

    0:07Shorts

📋 Key Takeaways
  • Choose between a one-time request, an ordered Workflow, an agent-prompt Shortcut, a direct-tool Shortcut, and a read-only research Automation; they are not interchangeable.
  • A meeting routine should inspect actual Do Not Disturb access and mode, record the prior state, propose Priority mode, and verify the result under configured approval policy.
  • When a saved routine changes, inspect its current actions, review complete parameters and dependencies, save the intended change, then start and verify the run yourself.
  • Unattended Automations support read-only web research, not scheduled phone writes. Recovery requires inspecting completed changes; stopping does not guarantee restoration.

Choose the Right Request or Saved Object

To automate multi-step tasks on Android, first decide whether you need help with the current situation or a reusable object. “Prepare for this meeting” is a one-time request. Saving ordered actions, saving an instruction for future planning, and scheduling research are different choices.

Keep the seven-stage control loop: intent, inspection, proposal, confirmation, execution, verification, and recovery. Intent defines the outcome. Inspection checks current state. Proposal names the change. Confirmation follows the configured approval policy. Execution performs supported actions. Verification checks what actually happened. Recovery identifies completed work and the remaining decision.

ObjectWhen to use itRequired definitionStart and result
One-time agent requestA task shaped by current phone stateYour intended outcome and relevant limitsYou start it; the agent plans supported steps
WorkflowA reusable ordered sequenceExplicit tool steps, arguments, and dependency orderYou start the saved sequence and verify its outcome
Agent-prompt ShortcutA reusable instruction requiring runtime planningConfirmed title, description, and intended behaviorYou start it; the agent interprets the saved intent
Direct-tool ShortcutKnown actions with complete parametersComplete, mapped tool steps and argumentsYou start the defined actions
AutomationUnattended read-only web researchResearch instructions and a local scheduleRuns on-device around the scheduled time, subject to Android delays

A prompt Shortcut is not a fully specified script, and a saved Workflow is not permission to run phone writes unattended. For fuller definitions, use FoneClaw Tools, Plugins, Skills, Workflows, and Shortcuts Explained.

Our configured model handles understanding and planning; enabled supported tools handle execution. AI Agent Phone Control on Android: Intent, Confirmation, Action explains that division. Completion means the intended result is verified, not merely that a plan was generated or saved.

Prepare a Meeting With Priority Do Not Disturb

Our official meeting Do Not Disturb demonstration shows an inspected, proposed, and approved Priority-mode scenario. It supports this meeting-preparation example, not a promise of autonomous scheduling or automatic restoration.

Begin with a bounded request: “Check the meeting I mean and my current Do Not Disturb state. Show a proposal for Priority mode before changing anything. Record the prior mode so I can review restoration afterward.” The steps below are proposed reader guidance, not a test we performed.

  1. Identify the meeting. Use the intended calendar event or a time window you provide. If several events fit, clarify which one matters before changing settings.
  2. Inspect Do Not Disturb. Read actual notification-policy access and the current interruption filter. Do not assume DND is off.
  3. Record the prior mode. Preserve the actual inspected value for later review. If it cannot be established, say so rather than inventing a baseline.
  4. Review the proposal. Name Priority mode explicitly. Do not silently change priority contacts, alarm exceptions, volume, or unrelated settings.
  5. Execute under actual policy. Complete any approval required by the configured mode and tool overrides.
  6. Read the result again. Confirm the active interruption filter rather than relying on a success sentence.

If DND access is missing or Android rejects the write, the supported change operation opens the relevant system access or settings screen. Handle that screen instead of immediately opening another copy. After the user grants necessary access, inspect state again before continuing.

After the meeting, restoration is a separate reviewed action. If the prior mode is known, you can request its restoration through the supported path and verify it. If it is unknown, choose the desired mode yourself. “Normal notifications” is not enough to establish a previous setting, and this routine does not guarantee a timer-based restore.

Review and Update a Reusable Routine

FoneClaw supports reviewing and updating saved Shortcuts when their actions change before saving and running them. Use the maintenance controls actually offered for the saved object; do not assume a particular menu sequence. Saving a change and executing the changed routine remain separate states.

Suppose a fictional “Meeting preparation” Shortcut contains a generic notification-setting instruction. You now want it to request Priority DND explicitly while retaining inspection and verification. First inspect the existing behavior. A familiar title does not tell you whether it is an agent-prompt Shortcut or a direct-tool Shortcut.

  1. Identify the object. Confirm which saved routine you are editing and what kind it is.
  2. Read the existing actions. Note what stays, what changes, and whether later steps depend on an earlier result.
  3. Define one bounded change. Replace the vague notification instruction with the intended Priority-mode behavior; do not add unrelated message or calendar actions.
  4. Review the complete definition. Confirm the prompt’s title, description, and behavior, or the direct sequence’s known tools and full arguments.
  5. Save the intended revision. Check that the saved object reflects the change before starting it.
  6. Start the routine yourself. Inspect current state and verify the outcome under the actual approval policy.

A Workflow stores an explicit ordered array of tool steps and their arguments. Dependencies matter: inspection must precede a proposal based on that inspection, and verification must follow the write. Do not treat saving the order as proof that a changing meeting or prior DND state is already known.

An agent-prompt Shortcut can preserve an instruction to inspect the current meeting and DND mode at runtime. A direct-tool Shortcut requires complete known mapped steps; it should not contain unresolved parameters disguised as a finished script. Stop and clarify rather than substituting guessed values.

For comparison with rule-building tools, Tasker Alternatives for Android: Free, Open-Source, and Voice Options covers that separate selection. Here, the maintenance question is whether the saved definition still expresses the task you intend to start.

Apply Actual Permissions and Approval Policy

Model capability, enabled tools, Android access, and approval settings are separate requirements. A model that understands “use Priority mode” does not gain notification-policy access by understanding the sentence. A saved Shortcut does not grant it either.

Approval follows the configured mode and per-tool overrides. Do not assume every read produces a popup or that every write uses an identical interaction. If you want to review the proposal before execution, explicitly request preparation only, then decide whether to continue.

Keep authorization scoped to the named action. Agreeing to a DND change does not approve sending a meeting message, modifying the event, or creating a personal ToDo. Drafting a message and sending it are different outcomes; opening Android settings and successfully changing a mode are different outcomes too.

An optional calendar or Memo write should be a separate user-started request with its own required access and approval behavior. For calendar work, confirm the selected calendar and device-local date and time, then inspect the actual returned times and saved event. A conversational “done” is not a destination record.

For voice-entry setup rather than routine design, use Android Voice Control: Voice Access Setup and Fixes. If the follow-up originates in recorded meeting notes, AI Audio Summarizer for Android: Record, Transcribe, and Review Notes explains source review before turning speech into an action.

Recover From the Actual Failed State

When a run stops, establish what completed before deciding what to repeat. The calendar inspection might have succeeded, Priority DND might already be active, and a later personal reminder might be missing. That is a partial result, not an untouched phone.

FailureInspectNext action
Missing DND policy accessAccess status and the system screen already opened by the change operationHandle the required access, then recheck; do not immediately launch a duplicate settings screen
Rejected or cancelled actionCurrent DND mode and whether any change took effectRespect cancellation or resolve the actual block before a new request
Meeting or phone state changedThe intended meeting and current interruption filterRefresh the proposal instead of applying assumptions from the earlier state
Uncertain completed writeCurrent DND state or the actual calendar/Memo destinationIdentify an existing result before retrying only the still-needed work

Report completed steps, blocked steps, and remaining work separately. If DND changed but a follow-up failed, say that DND is active and the follow-up remains unresolved. Do not rerun a whole sequence merely because its last step did not finish.

Restoration also requires fresh inspection. The user may have changed notification settings after the routine ran. Compare current state with the recorded prior mode and ask whether restoring that mode is still desired. Do not overwrite a deliberate later choice in the name of recovery.

Where supported, a user-requested restoration can return to the known earlier mode, followed by another status check. Without a reliable prior value, ask for the desired state instead of assuming DND should be off. Stopping execution does not roll back a completed write, and not every side effect has an automatic undo.

For uncertain calendar or Memo creation, inspect the saved destination before repeating creation. Correct an existing record through the supported path when appropriate. Recovery should resolve the specific failed boundary, not introduce duplicate items or discard successful work.

Separate Phone Routines From Scheduled Research

User-started phone routines and unattended Automations have different execution scopes. The meeting DND routine is something you start when needed, with current state, access, and approval considered. Unattended scheduled tasks are restricted to read-only web research; they do not schedule DND changes, SMS sending, or calendar and Memo writes.

A suitable fictional Automation could prepare a weekday brief about publicly available transit disruptions. Specify the research scope and local schedule. It runs on-device around that time, but Android may delay it. Treat the result as research to review, not a guaranteed punctual phone action.

You can search existing scheduled tasks and inspect a pending Run Now request. Search before creating another task with the same purpose. Run Now is deduplicated against the same day’s scheduled execution, with at most one execution that day; it is not a guarantee of a second run after the scheduled task already completed.

A research result or notification does not authorize a follow-up write. If you want to save a personal ToDo or change a calendar entry afterward, start that separate supported task and verify its destination.

Before reusing the meeting routine, check five things: the correct saved object, reviewed changed actions, current DND access and mode, the verified result, and the restoration choice. Our supported capability areas are described on FoneClaw Features. The practical benefit is a repeatable, inspectable sequence—not unattended authority over every phone setting or a promise that every interruption will restore itself.

Try FoneClaw