Check the input, then compare the complete deliverables
Explore the original materials, item-by-item outcomes and full report below. Attempt the input independently before using the reference to find omissions or disagreements.
These are text samples. No audio, speech recognition, photos or mobile accounts were accessed or tested.
Original synthetic material. Try the input independently, then compare the reference results.
Original synthetic input
{
"synthetic": true,
"case": "mobile-voice",
"data": {
"recorded_date": "2026-10-17",
"timezone": "Asia/Shanghai",
"requests": [
{
"id": "V01",
"transcript": "Register this receipt. The amount might be two hundred eighty, or two thousand eight hundred.",
"attachment": "No image supplied"
},
{
"id": "V02",
"transcript": "Next Monday’s meeting moved to 3 pm. Redo the brief; do not touch Calendar.",
"target_event_id": null
},
{
"id": "V03",
"transcript": "List today’s emails that need my reply.",
"mailbox": null,
"query": null
}
]
}
}Processing rules
Retain the transcript verbatim, then restate action, object, scope and gaps. Do not write an ambiguous amount or unidentified event. Recreating a brief is not changing Calendar. Voice does not skip review. Return scoped results to the current private conversation.
Complete reference results under this tutorial’s rules; nothing was written to a third-party service.Scroll horizontally within the table if needed.
| Request | Known intent | Ambiguity or gap | Current output |
|---|---|---|---|
| V01 | Organize a receipt | 280/2800; image and target sheet missing | Review only; do not register an amount |
| V02 | Rebuild meeting brief | Event id missing | October 19 at 15:00 needs verification; no Calendar edit |
| V03 | List emails needing reply | Mailbox, time window and query unknown | Ask scope before reading |
V01 needs the original image, confirmed amount and target sheet. No entry has been created.
V02 asks to regenerate a brief; it does not authorize rescheduling or notifying attendees.
Voice task confirmation — text exercise V01: receipt intent; two possible amounts; image missing. Await confirmation; no sheet write. V02: with 2026-10-17 in Asia/Shanghai, next Monday is October 19; map 15:00 to a specific event before updating the brief. No Calendar edit. V03: identify mailbox, today’s time bounds and limit before a read-only list. No recording, OCR, account reads or sending.
Prepare the materials and check your tools
Start with the practice files
- Use the three text transcripts without a microphone.
- No image is provided; do not claim OCR.
- Check amount, event and mailbox gaps in V01–V03.
Before using real services
- Open the same Bot conversation on mobile; distinguish dictation from live voice.
- Review transcription and attachments before sending; complete login yourself.
- Return to desktop for configuration or validation unavailable on mobile.
Complete one file-based exercise first. For real services, verify the actual account, available tools and operation scope, then follow the acceptance check at each stage.
Download the original input and rules ↓Execution and takeover boundaries
The actions below require the agreed inputs, accounts and scope. Configuring a Routine does not expand that authorization.
Can run automatically
Process supplied, confirmed text, transcripts or photo descriptions into a reviewable task draft.
Requires confirmation
Confirm unclear names, amounts and dates. Recording, calls, messages, private uploads and external tools need separate approval.
Take over after a failure
Stop at a draft when transcription or image context is incomplete. Show the original and ambiguities for user correction or desktop handling.
Check dictation, voice and conversation context
Dictation prepares a message; live voice is a continuing conversation. Verify the Bot, transcript and attachments before relying on the context.
Read only the three synthetic transcripts. List ids, exact wording, attachment state, current date and time zone. Do not request microphone access or read photos/accounts. Return an input inventory.Traceable transcripts and missing attachments.
V01 still has no image or OCR result.
Restate action, object and scope
Speech often omits accounts and dates. Make a task card first, and restate concrete relative-date interpretations for review.
Create V01–V03 task cards with action, object, account, time, limit, allowed operation and gaps. Separate redo brief from edit Calendar. Show exact relative dates and the date/time-zone basis.Three task cards.
V02 permits only a new brief, not event changes.
Ask only the questions that resolve the ambiguity
A mistaken amount or same-named event can change the result. Ask about the missing field rather than making the user repeat everything.
For V01 ask amount, original image and target sheet; V02 event id; V03 mailbox and today’s time zone/window. Separate work possible now from work awaiting answers. Infer no amount or credentials.A minimal clarification list.
Neither 280 nor 2800 is chosen; no unidentified event is changed.
Prepare a draft before executing clear actions
Existing text can be organized while a dependent operation waits. Understanding is not completion.
Generate a confirmation note from the synthetic input only, listing questions and unexecuted actions. Do not connect email, Calendar or Sheets. Keep unanswered items pending; do not mark them complete.A complete confirmation note.
All three retain their gaps; no registered amount or rescheduled event is claimed.
Review actual results in the same mobile conversation
A success message should correspond to an actual source record. Let the user handle login or verification steps.
Choose one clearly scoped real read-only request. Restate account, range and goal; return results, source ids, coverage and unexecuted items. If tools or permissions are missing, deliver text and the blocker without changing access paths.A traceable small trial or blocker.
Results are verifiable; oral confirmation does not hide failure or non-execution.
Template the parameters that are easy to mishear
Reusable fields help but money, dates and identities still require each-run confirmation. Mobile may not support every desktop setup step.
Save a mobile-task-intake template: original wording, restatement, time zone, object id, gaps, allowed actions and result review. Ask about money/time ambiguity. Do not create a Routine; identify desktop handoff where needed and send no third-party notifications.A voice intake template.
The template keeps restatement and object checks on every run.
Find the evidence before choosing the next action
Three o’clock became thirty
Confirm a concrete valid time from the transcript; do not silently correct it.
The desktop chat is missing on mobile
Check account and Bot before submitting a duplicate task.
The photo amount is still unclear
Keep it unknown and inspect the original; this page guarantees no OCR accuracy.
The task continued without a result
Inspect the same conversation and status; distinguish progress, blockers and failures before resubmitting writes.
Check every item before finishing
These page-local checkboxes reset on refresh. Checking them does not verify a real service connection.
Checked 0 / 6 items
Download the inputs separately from the reference answers
Start with input and rules without showing the reference answers. After acceptance, save the report and complete prompt set.
source-data.json↓Rules and scopeDefinitions, exceptions and delivery requirements.workflow-rules.md↓Item-by-item reference resultsCompare every input with its outcome.expected-results.csv↓Complete reference reportResults, evidence and missing information.reference-report.md↓All stage prompts6 instructions with expected outputs and acceptance checks.practice-prompts.txt↓Check documented capabilities and exercise results separately
Content updated: · Editing a page does not reverify its sources or live workflow.
Documentation review: Unverified · no review date recorded
No case-specific documentation review is recorded. References below are provided for review and do not establish verification.
Live workflow verification: Unverified
Live accounts, external writes, scheduled execution and long-term operation are unverified. Synthetic exercises and page checkboxes do not establish live verification.
These are text samples. No audio, speech recognition, photos or mobile accounts were accessed or tested.
- Grok Bot: mobile ↗
Compare dictation, voice, attachments and mobile/desktop capabilities; no recording was tested here.
- Grok Bot: apps and the shared computer ↗
Check connected accounts, available tools, visibility and user takeover.
- Grok Bot: Skills and Routines ↗
Validate a manual run before saving a method or scheduling it.
Try another small batch of your own material
Retain original input and operation records. Verify that each result can be traced before adding connections or recurring runs.