---
name: tutorial-verified-done
description: Make a user-facing feature tutorial-verified done through narrated real-browser journeys, one-second visual evidence, exact-byte independent verdicts, and a recoverable delivery loop. Use for UI completion gates, tutorial production, visual proof packages, and publish-or-feedback handoffs.
version: 1.0.0
updated: 2026-08-14
---

# Tutorial-Verified Done

Treat the tutorial as the release artifact for every completed five-phase build, not documentation added later. Map every phase-0 user-story step one-to-one to a sentence for a layperson with zero technical detail, a visible human action, and a visible outcome. Preserve stable requirement IDs back to the raw prompt, transcript, or microphone submission; a missing requirement fails the gate.

The same build wave owns this proof. Before reporting, re-read the original human prompt, phase-0
story, and final specification; run the E2E capture and regressions; self-score every story and final-spec step against named evidence; and
write `docs/evidence/receipts/<item-id>.md`. Use the one canonical
[evidence receipt schema](references/evidence-receipt-schema.md) and its machine checker; a report
without a valid receipt is `FAIL`.

Run the complete journey in production-identical desktop light, desktop dark, mobile light, and mobile dark variants. Capture at least one frame per second from landing through success, record narration, derive the transcript from that audio, and assemble a labeled contact sheet for each variant. Declare which reference controls layout and which controls color and tone. Describe the visible frames supporting every claim.

Any error, empty state, dead end, stale surface, visual divergence, failed variant, or transcript-to-frame mismatch invalidates that run. Repair the product and re-record the whole affected flow. A validator, manifest, consumer route, historical receipt, scripted text timeline, or loose media file is contract evidence only; it is not an executed proof package.

The keeper relaunches dead work only for its remainder, re-captures before unrelated work after a
capture failure, and opens a bounded automatic rework wave after a missing receipt or twin `FAIL`.
Only exact external boundaries go to the human: publish clicks, physical devices, credentials,
current spend authority, production mutation, outbound communication, destructive action, or human
values decisions.

Pin every proof package and verdict to the exact reviewed bytes or commit. Two independent clean-context max-effort reviewers inspect the raw request, running code, four contact sheets, audio-derived transcripts, checksums, and served-artifact identity. Each returns binary `PASS` or `FAIL`. A conditional finding names the exact edit and requires new evidence plus a fresh verdict over the new bytes. Only unanimous verdicts over identical bytes may pass; the implementer never self-greens.

Attach a passing tutorial idempotently through the progress-tracker API. With current messaging authority, send the user a push notification containing the tutorial. After the user watches, offer a one-click publish control when the agent has publish permission or a linked microphone-feedback control; never silently publish or send.

Evidence ceiling: Gallery contains useful four-variant browser capture, one-second frames, contact sheets, inventories, and validators, but the surveyed tree did not prove the complete narrated, reviewed, uploaded, READY, notified, watched, and publish-or-feedback loop. Do not represent that partial prior art as finished.