Question Lab / Human direction, mid-task
Sometimes the useful next step is a choice only the person can make. A question in native chat can pause the work, receive that direction and let the same task carry on.
Actual native execution / recorded replay
The film shows the real native flow. These controls move through its recorded moments; they make no new model request.
Isolated iPhone Simulator. Automated test participant. A local API shim forwards real Claude Engine inference. This build includes an unreleased Claude-only answer-key correction; it does not demonstrate the live PanelForge API.
The fictional Lantern Makers workshop
01 / 04The fictional workshop needs a check-in. The task asks Claude to pause for a direction before drafting.
Plan a check-in for the fictional Lantern Makers workshop. Before drafting, ask whether to prioritize Quick setup or More customization. Use my answer to give three short practical steps.
Claude calls its built-in question tool. The native card shows two choices while the original invocation remains open.
Should I prioritize Quick setup or More customization for the Lantern Makers check-in?
The replay follows the recorded Quick setup answer. More customization has no recorded result here.
The automated participant selects Quick setup and sends the native answer. The original pending question receives it.
The request reached the original pending question. Its request and stream identities matched.
The actual tool result confirms Quick setup. That same invocation resumes and drafts the short plan below.
Chosen direction: Quick setup
Actual model text. This is a plan; none of the actions in it ran.
One original task. The answer has not been chosen.
Authentic native pixels
Details from the original device screenshots, paired with the recorded story. Each link opens the full-size capture.
The original invocation pauses with a real built-in question.
Isolated Simulator; automated participant; local API shim; unreleased Claude-only correction.
Full-size original
The automated participant selects Quick setup. Only this choice was recorded.
Isolated Simulator; automated participant; local API shim; unreleased Claude-only correction.
Full-size original
The final text names the answer and keeps the plan short.
Isolated Simulator; automated participant; local API shim; unreleased Claude-only correction.
Full-size originalThe previews crop the original screenshots for readability. They do not recreate the native screen. The final text is a drafted plan; none of its actions ran.
Follow the causal receipt
The original request stayed open. The native answer resolved its pending question, and the actual tool result confirmed the choice before the final text followed.
Inspect the public receiptActual built-in tool result
Your questions have been answered: "Should I prioritize Quick setup or More customization for the Lantern Makers check-in?"="Quick setup". You can now continue with these answers in mind.
The question, native answer, pending future, tool result and final completion were checked against matching private identities. The public receipt omits those opaque IDs.
The visible card’s selection is under the UI header Priority. The native request sent the answer under the full question text, as the built-in Claude tool requires.
{
"action": "answer",
"answers": {
"Should I prioritize Quick setup or More customization for the Lantern Makers check-in?": "Quick setup"
}
}This is the captured native payload with its opaque request ID removed, verified against the private request. The shim forwarded it unchanged.
The existing SwiftUI chat card, controller, SSE decoder and response API ran in a separate iPhone 17 Pro Simulator build. A minimal local API shim forwarded original SSE and the native answer to the original Claude Engine handlers. The model used its real built-in question tool through subscription inference.
The final native run passed its exact answer-map regression and actual UI flow: two tests, no failures. The successful run made one question tool call, then produced 25 text events after the answer and a 232-character final plan.
The fictional task asks for a workshop check-in plan after choosing between Quick setup and More customization. Claude asks: “Should I prioritize Quick setup or More customization for the Lantern Makers check-in?” The automated participant selects Quick setup and sends it. The actual tool confirms that answer. The same invocation resumes and visibly produces Chosen direction: Quick setup, followed by a short invite, a shared document and a 20-minute session. Those are drafted suggestions; no invite, document or session was created.
The first actual run returned HTTP “answered,” but the built-in tool said the questions were unanswered. The native card had keyed its map by the short UI header. An expected failing regression exposed that mismatch; the isolated Claude path was corrected to use full question text.
This correction is unreleased. The capture uses an isolated Simulator and local API shim; it does not validate deployed Laravel persistence, production routing or the unmodified released native app. Codex, spoken answers, macOS cards, multi-select and reconnect behavior were not tested here.
bf89abc1646bd12133a0f22bc2d51f328681325048d380467d270c6f5fb10cbfd68923985a973c3bDifferent ways a person can intervene
This is Claude’s built-in AskUserQuestion: a task needs an answer, and its original question call waits for one.