Voice Follow-Up: Check the Reference and Draft | Yihu
Before saying continue or change it, check the referenced material, current conversation and existing draft. Try Yihu’s no-mic sample, then review privacy and App Store setup.
Preview privacy before you try
The linked preview uses preset material, needs no login or microphone, and keeps edits only in the open page. Reloading resets edits. Optional read-aloud uses the browser’s voice service; usage analytics require consent.
Keep private originals in approved tools. In the app, on-device recognition and the destination’s storage, sync and access rules are separate steps. Review the installed version before using a source-described workflow. Read the data-flow guide
US listing verified; availability depends on your storefront. No Mac or self-signing for store installation. Bring your workspace address and account; the optional speech model is about 950 MB.
Name the reference, then check the conversation and draft
Name the passage or item instead of relying on it or previous. Check the current workspace, actual service conversation and unsent target draft, then inspect the history the service supplies. Yihu has no universal Agent backend or cross-workbench memory. If the earlier result is unclear, keep the text and inspect the destination instead of resending to restore context.
Different from setup, correction and repeat delivery
Readiness checks prerequisites; correction checks text copies. Here the next utterance needs an explicit antecedent and authorized target. Do not automatically resend the previous passage to restore context, or treat visible page text as proof the service read/executed it.
Follow-up context preflight
| Check | Material to establish | Synthetic example | Do not infer |
|---|---|---|---|
| Referent | Which previous phrase/item is meant | Synthetic phrase A, not a real order | It points to the correct business object automatically |
| Workbench | Current authorized HTTPS source | Existing own workbench only | Reachability is server identity/write permission |
| Page/conversation | Current page and actual service session | Is this the reviewed sample-draft session? | Navigation restoration is model memory restoration |
| Composer | Editable unsent field on the original page | Focus the compatible draft first | Arbitrary editor/terminal insertion is permitted |
| History scope | Service actually provides the required context | Visible history still needs scope/content review | Visible history necessarily reached the Agent context |
| Intended operation | Organization, proposal and execution differ | Edit sample wording only; do not send | Follow-up tone grants execution authority |
| Prior handoff | Inserted/staged/uncertain state | Inspect an existing draft | Missing receipt permits another delivery of prior text |
| Changed destination | Captured and current page/state | Recheck guards after navigation | A source switch automatically retargets an opened note |
A less ambiguous synthetic draft
Ambiguous: Continue; change it to two items.
More explicit: Review only the wording of local synthetic phrase A. Item count is unconfirmed; do not send or execute.
Still check: The current workbench/conversation is the authorized sample destination.
This is wording guidance, not guaranteed model interpretation or an Agent/service call. Do not copy a private conversation/customer data into a public example to add context.
Destination binding is not backend context
Native draft delivery has source/page/request/host-refusal guards. Explicit note insertion uses the workbench/page captured when notes opened. These constrain where text may go; they do not determine which history a model receives, how pronouns are interpreted or whether a service persists conversations. A passed target guard does not establish correct understanding.
Implemented WebKit interaction restoration concerns navigation scope, not cross-origin identity copying or a unified conversation database. With unknown source, prior result or conversation, retain text and stop. Do not guess context by changing pages, repeating old rounds or clearing markers.
FAQ
Does Yihu remember all prior speech across workbenches?
No such cross-workbench memory is asserted. It is a browser shell with voice drafts; the actual Agent conversation belongs to each service.
Does restored page history establish restored model context?
No. Navigation/scroll state and the history supplied to a backend model are different facts requiring actual service evidence.
Can continue or previous omit the referenced object?
Avoid that when the object is unclear. Check material/conversation and state a synthetic reference within authorized scope rather than asking a reader to infer a real target.
Can a follow-up automatically resend the prior round to restore context?
No blind resend follows. It may already be delivered or uncertain; inspect destination content and retain original request/one-handoff boundaries.
Does explicit note insertion follow page changes after opening notes?
Do not assume so. The note target is captured on opening and checked before insertion. Follow existing guards without retargeting by bypass.
Does apparent understanding authorize every requested operation?
No. Interpretation, explicit user intent and server authority differ. Original confirmation, write permission and receipts remain necessary.
Reading and source
Voice-draft choices · Web output/capture transition · Originals and note targets · Manual summary provenance · Start · Privacy
Evidence: draft-target guards, note target capture/checks, README service limits.
