tiktok-gemini-video-workflow

TikTok Gemini Video Workflow

Goal

Convert “the user only submits product images and viral videos” into a recoverable, traceable production pipeline. Always use Feishu Base as the task state source, and use real downloaded files and read-back results as completion evidence.

Capabilities that must be used in combination

When involving Feishu Base, first use lark-base, and through lark-cli base +... read the real tables, fields, and records.
When involving nine-grid generation, use imagegen; clearly write the role of each input image.
When involving Gemini or Flow web pages, use chrome:control-chrome, reusing the Chrome that the user explicitly selected and is logged into.
When involving single-account, multi-account switching, quota polling, or account logs, read account-routing.md。
When involving scheduled polling or continuing after quota recovery, first look for the automation tool provided by the product; do not use shell cron instead.
When involving visible star watermarks, prioritize using the installedgemini-watermark-remover Skill; if missing, follow setup.md to obtain permission before installing.

Default entry

Gemini professional video entry:https://gemini.google.com/videos
Google Flow：https://labs.google/fx/tools/flow/
Base and Flow projects: use the links provided by the current user; do not hard-code any project, table, account, or record ID in the Skill.

0. Environment pre-check

According to setup.md check Node.js, official@larksuite/cli, Feishu user authorization, Chrome connection,ffmpeg/ffprobeandgemini-watermark-remover。
lark-cliIf not installed, first explain what will be installed and obtain the user's consent, then install and configure; if already installed, do not repeat the operation.
Only start processing tasks after Feishu read-only access, Chrome login state, local directories, and post-processing dependencies are all successfully verified.

Confirm input and output before starting

At the start of each new session or new batch, first confirm the following with the user all at once; this is a hard startup threshold:

Task table: whether the user already has a Feishu Base for this workflow.
- Yes: ask the user for the Base link, table/view, and processing scope.
- No: confirm the new Base name and data table name with the user, then create them according to “Initialize task table” below.
Working directory: which directory to place intermediate files such as downloaded materials, frame extraction, transcription, nine-grid, and tail frames in.
Finished product directory: which directory to place the original downloaded videos, watermark-removed versions, and final confirmed versions in.
Browser and account: which Chrome profile and which Gemini/Flow account to use; set an alias for the account that does not contain private credentials.
- Single account: confirm the account alias, Chrome profile name, and allowed entry.
- Multiple accounts: further confirm enabled accounts, priority, whether automatic switching is allowed, and whether the same task must remain on the same account; according to account-routing.md establish account registration.
Delivery method: upload to Feishu immediately after generation and enter待验收, or save only locally first and upload after the user confirms all together.
Notification method: only update the Base, or also send a Feishu message when the status changes; sending messages requires the user to provide a resolvable individual or group chat target and authorize its uselark-im。
Scheduling method: run manually this time, or set up scheduled polling; when not set, never create background tasks.
Submission form version: use the basic version that only collects product images, viral videos, and an optional product name, or add optional items such as target country/language, product selling points, prohibited content, and narration/music preferences. When there is no explicit choice, ask first and do not expand the form on the user's behalf.

If the user has already explicitly provided it in the current session, reuse it directly and do not ask again; do not default to the old directory or old browser configuration across sessions. Do not start downloading, generating, or submitting quota tasks before the working directory and finished product directory are confirmed.

When the user chooses manual operation, only process tasks within the scope confirmed for this run; after the batch ends or a waiting condition is encountered, write back the status, quota log, and next step and exit, without creating scheduled tasks, resident background processes, or automatic continuation.

After confirmation, use an independent working subdirectory for each record, suggested as<工作目录>/<record_id>/; finished products uniformly go into the finished product directory. Do not scatter finished products in Downloads or temporary directories. Immediately after the browser download completes, move it and name it according to产品名-视频序号-入口-处理状态-画幅-时长.

1. Initialize the task table

The user already has a Base

Use the URL provided by the user to executelark-cli base +url-resolve --as user, obtain the realbase_token。
Read the target table, fields, views, and forms; do not treat the Wiki token as a Base token.
Compare against base-contract.md to perform a structure check.
When fields or status options are missing, list the differences and complete them after obtaining user authorization; this Skill is fixed at a maximum of two segments and does not create segment 3 or 4 fields.
After the structure check is complete, ask about the working directory, finished product directory, browser, and delivery method.

The user does not have a Base

Ask for the new Base name and initial data table name; if the user has no special naming, you may suggestTikTok爆款视频复刻and视频任务, but must not decide silently.
Uselark-cli base +base-createto create the Base, and create all fields defined by base-contract.md in one go. After creation, read back the Base, tables, and fields.
Before creating the form, first ask the user to choose the basic version or the extended version.
- The basic version only displays:产品图(required attachment),爆款视频(required attachment),产品名或备注（可选）(optional text).
- In addition to the basic version, the extended version adds optional items such as target country/language, core product selling points, prohibited content, and narration/music preferences according to the user's choice; do not create all of them by default. Use the system creation time for the date and do not require the user to fill it in.
Create待处理队列、处理中、待验收and已完成views, and set filters according to task status.
Return accessible Base and form information; only after confirming the user can open them, proceed to task processing.
New creation, field creation, form creation, and view creation all uselark-base; do not merely generate a field list or local template to pretend it has been created.

Core workflow

1. Handle manual review first, then claim new tasks

Read the real field structure as well as验收结果、验收意见and candidate records.
First handle manual review results:
- 通过: change任务状态to已完成and read back; keep the review result as evidence.
- 废弃: change任务状态to已废弃and read back; subsequent runs will no longer process it.
- 返工: takes priority over all new tasks. Read验收意见to determine the segments that need to be redone; first change the status to the corresponding processing stage, then continue from that segment. When redoing segment 1 and segment 2 exists, segment 2 and the continuity assets must be regenerated together with the new segment 1.
For rework tasks, only after the new finished product is uploaded and successfully read back should you clear验收结果, set the status back to待验收; do not clear before generation. Append this acceptance feedback, the reworked segments, and the results to运行日志, keep验收意见for traceability.
After there are no rework tasks, select only new records that simultaneously satisfy the following conditions:
- 产品图is not empty;
- 爆款视频is not empty;
- 任务状态is empty or is待处理。
Process by creation date from earliest to latest.
Before downloading and analysis, first change the record to拆解中, using this status as the task lock to avoid duplicate execution.
When claiming, if任务IDis empty, write the unique task ID; at the same time set当前片段to 1, record最后成功步骤as已认领。
Skip拆解中、分镜图生成中and视频生成中records already claimed by other running instances.待用户操作、等待额度恢复and待验收are handled according to the corresponding recovery rules;已完成and已废弃are no longer processed;失败is only re-claimed when the user requests a retry or the retry policy explicitly allows it.
Before resuming, read当前片段、最后成功步骤, the planned segment count, existing storyboard images, tail frames, and finished product attachments, and continue from the last successfully read-back step. Do not guess progress based only on local file names, and do not redo steps that have already been accepted and uploaded.

For read and write rules, see base-contract.md。

2. Download and break down the viral hit

Download the product images and the viral hit video to an independent working directory.
Use ffprobe to read the duration, resolution, frame rate, and audio track.
Extract frames every second or generate a contact sheet to identify: the first-frame hook, action beats, scene changes, product exposure, transition points, and ending action.
Inspect on-screen text frame by frame and transcribe it verbatim. Distinguish text hooks, ordinary descriptions, product packaging text, and platform watermarks; record the original text, capitalization, spelling, punctuation, line breaks, position, appearance interval, and style. Do not correct grammar or rewrite copy on behalf of the original author without the user's request.
When there is dialogue, prioritize transcribing the audio track; distinguish visual highlights, text highlights, dialogue highlights, and sound highlights. Determine whether the narrative depends on "text + scene/action" working together, and do not mistakenly break down a text-reversal video into a pure product demonstration.
Do not copy solutions unrelated to the product. Preserve the viral hit structure and change the actions and selling points into content that the current product can deliver; the original text that carries the core hook must enter the new script and post-production checklist.

3. Decide the number of segments

Supports 1–2 segments. First decide the default number of segments based on the original viral hit's total duration, then fine-tune according to action and transition boundaries; do not mechanically cut in the middle of an action.

When the original viral hit's total duration does not exceed 15 seconds, uniformly make 1 segment, compressing the complete highlight into the single-segment duration that the current professional entry point can generate.
If the original viral video's total duration exceeds 15 seconds but does not exceed 30 seconds, make 2 segments.
If the original viral video's total duration exceeds 30 seconds, stop and ask the user whether to compress it into two segments or switch to a shorter reference video; do not create a 3rd segment, truncate the latter half, or silently lose content.
For Gemini, the actual single-segment length is subject to the current professional video entry point; prompt duration cannot override system settings.
Flow default setting is Video / 9:16 / 1x / 10s, but each time the live interface prevails.
The 2nd segment must use the real last frame of the 1st segment's downloaded final output as the continuity input.

4. Uniformly use a 3x3 grid

All clips uniformly use one complete 9:16 3x3 grid storyboard; no longer choose between keyframes and a 3x3 grid.
The nine cells must express the continuous action of the same clip from left to right and top to bottom; the same character, face, hairstyle, clothing, scene, lighting, and product identity must remain consistent.
The 3x3 grid is always uploaded as one complete storyboard image; do not split it into nine reference images.
Last-frame continuation is an independent hard rule: the 2nd segment's 3x3 grid cannot be generated in advance. The 1st segment's real final output must first be completed and downloaded, then the real last frame extracted; use第1段真实尾帧 + 第1段九宫格 + 产品图to generate the 2nd segment's 3x3 grid, and have the 1st cell accurately inherit the real last frame.

5. Generate and immediately return the storyboard image

First write back爆款拆解、新脚本、分段数量and each segment's prompt, then change the status to分镜图生成中。
A single-segment task generates a unique 9:16 3x3 grid storyboard; a multi-segment task only generates the 9:16 3x3 grid for the segment currently pending production first, and does not generate subsequent storyboards all at once in advance.
Preserve product color, structure, packaging, Logo, and character identity; avoid extraneous text, watermarks, extra limbs, and collage leakage. Do not draw the original work's necessary text hook into the 3x3 grid; record it separately as a post-production subtitle asset to avoid garbled text from image-to-video.
Each time a storyboard image is generated and accepted, immediately upload the corresponding分镜图 ifield for that record; do not wait until all storyboard images are complete to upload them in batches.
After uploading, immediately read back the attachment field to confirm the file exists, then set当前片段=i、最后成功步骤=分镜图 i 已上传。
For a single-segment task, after the storyboard and video prompt are complete, change the status to视频生成中. For a two-segment task, at分镜图1immediately enter video 1 after acceptance and upload, without waiting for storyboard image 2.
The 2nd segment's storyboard can only be generated after the 1st segment's video has been downloaded and the real last frame extracted. Use第1段真实尾帧as the primary continuity reference and the 1st cell of this segment's 3x3 grid, use第1段九宫格to lock the character, hairstyle, clothing, scene, and camera style, and use the product image to lock the product; only change the actions and product states in the subsequent 8 cells, and do not randomly regenerate the character.

Before writing any storyboard image or video prompt, you must fully read prompt-templates.md. All prompts must uniformly use English; only product names, entity labels/Logos, original on-screen text, and dialogue that must be preserved verbatim may remain in their original language. Do not merely translate old Chinese prompts directly into English; you must fill in the template with input responsibilities, reference priorities, locked items, cell-by-cell/second-by-second actions, realism, action physics, audio, text post-production, and negative constraints.

Each 3x3 grid prompt must be written in English structure:Inputs / Goal / Story / Reference Roles / Character & Product Lock / Scene Lock / Visual Style / Storyboard Layout / Panel 1–9 / Continuity Requirements / Text & Logo Rules / Negative Rules. Each cell must include time, shot size, narrative function, visible action and state, camera movement, and emotion; the 3x3 grid remains one complete 9:16, 3×3 storyboard image, and is not changed into a six-grid or nine independent reference images.

Storyboard image prompts and video prompts are two independent outputs and must not be mixed in counting. The storyboard image is fixed to generate one complete 9:16 nine-grid, expressing only 9 continuous key states; the video prompt must re-read and break down the reference video, writing each identified frame state separately asVIDEO FRAME 01...N. The number of frames in the video prompt is not limited by the nine-grid: if 16 frame states are broken out from the original footage, 16 entries must be written, and they must not be compressed to 6, 7, or 9 entries to match the nine-grid or an assumed shot limit.

Each video prompt must be written in the English structure:Inputs / Goal / Reference Roles / Output Format / Global Fixed Requirements / Character & Product Lock / Scene Lock / Camera Plan / Complete Video Frame List / Realism Pass / Action Physics Pass / Audio / Text & Logo Rules / Continuity Requirements / Negative Rules。Complete Video Frame ListEach frame state in it must be a separate block and fully filled in:Timecode and Duration / Reference Images / Continuity In / Shot Purpose / Framing and Camera / Character and Limb Actions / Product and Part State / Functional Interaction / Secondary Characters / Environment and Lighting / Audio / Transition In & Out / Continuity Out. Multiple different frame states must not be compressed into one paragraph.

Each action must specify the character's own left hand, right hand, left foot, or right foot; product operations must specify the holding part, working end, contact object, movement direction, resistance, andbefore → action → afterstate changes; paired items and detachable parts must maintain conservation of quantity, assembly state, and position. The sum of the durations of all video frame entries must exactly equal the real-time generation duration of the professional entrance, and time codes must not overlap, leave gaps, or omit any frame state from the original footage. Fast-cut frames may be shorter than the average rhythm and must not be deleted merely because their duration is short; when two segments are needed, all frame states must be allocated according to this Skill's established total-duration rules, and segment 2 must continue from segment 1's actual tail frame.

Character identity must be locked by the explicitly specified main model reference image, and the main model's specific facial features must not be re-described or guessed in the prompt; secondary characters must not reuse the main model's face or each other's faces, and identity, age range, body type, clothing, and role must remain stable across shots. Multiple product images jointly constrain different views of the same product and parts, and the product must not be duplicated according to the number of images. The prompt must also first choose a single visible text strategy:NO VISIBLE TEXTWhen this is used, subtitles, titles, time codes, brand names, Logos, packaging text, account names, platform names, UI, and watermarks are prohibited;PHYSICAL LABELS ONLYWhen this is used, only the real physical labels/Logos in the product reference images are retained, and overlay text and model-generated subtitles are still prohibited. The two strategies must not be used at the same time.

Short prompts that write only a one-sentence plot summary, vague actions, do not cover the full duration, lack a frame end state, merge multiple frame states into one paragraph, or cannot carry over to the next frame are unqualified; when the mouth is not in frame, lip sync must not be required and narration should be used instead. When the user requests multiple languages, by default generate an independent prompt/final video version for each language, unless the user explicitly requests bilingual content in the same video.

When the original work has on-screen text, the prompt must also stateTEXT HOOKresponsibility: explain how this text, together with the characters, scene, and actions, completes the narrative. Do not rely on Gemini/Flow to correctly generate text; the prompt requires the model not to render subtitles, and the exact text, typography, and time range must be placed in the post-production checklist. After generation is complete, check and add text according to postprocess-sop.md .

The output specification must clearly stateGoal、Output FormatandNegative Rulesin the three places VERTICAL PORTRAIT, aspect ratio 9:16 (width:height), 1080×1920 orientation, and declare that this is the frame aspect ratio rather than duration/timestamp, and prohibit 16:9 landscape. The prompt cannot override Gemini's actual generator frame and duration settings; the professional video entrance must be used and the aspect ratio control must actually be set toPortrait/肖像（9:16）, and submission is only allowed after reading back the selected state. Final acceptance is based only on the real width and height of the downloaded file, and the width must be less than the height.

Each Gemini task defaults from the leftVideos/视频Enter the professional video page: upload the storyboard image and product image for this segment; for a continuous segment, also add the real last frame of the previous segment, paste the complete prompt, actually select 9:16, then submit. Multiple reference images must be uploaded one by one and you must wait for each to finish; do not select multiple at once. The appearance of a local thumbnail does not mean the attachment has been submitted with the request. Do not rely on attachments, aspect ratio, or context left over from ordinary chat or a previous task. When the professional entry cannot reliably select 9:16, stop consuming Gemini quota and switch to Flow. For the specific tested recovery steps, see browser-sops.md。

6. Choose the generation entry based on quota

Fixed priority:

First use the dynamic quota of the Gemini professional video entry. With multiple accounts, first rotate through all available Gemini accounts, then enter the Flow stage; for account ordering, stickiness, and switching rules, follow account-routing.md。
update after each successful submission or quota check当前生成入口、Gemini额度状态、Flow额度状态、额度检查时间and运行日志。
After the same account has successfully submitted 3 Gemini videos in a row, mark only that account as疑似耗尽（3次经验阈值）, do not waste a 4th request just to test; with multiple accounts, switch to the next available Gemini account. This threshold is only responsible for switching early and is not equivalent to the platform's official limit.
After Gemini reaches its limit, read设置 → 使用限制for the current usage, short-cycle recovery time, and weekly quota recovery time, and record them. If the page does not indicate a date or time zone, keep the original text and do not convert it yourself.
If there are still pending segments, switch to Flow immediately; Flow credits and Gemini usage limits are managed separately.
When Flow credits are insufficient or used up for this round, stop using Flow and immediately recheck the Gemini usage limit; if Gemini has recovered, switch back to Gemini to continue the unfinished segments, and do not carry over the old quota conclusion from before entering Flow.
Only after rechecking, if Gemini is still restricted and Flow is also unavailable, switch to等待额度恢复. Purchasing or automatically topping up AI Credits is a paid action and must obtain user confirmation before acting; only write下次运行时间if the exact recovery time can be parsed, otherwise record the original page text and wait for the user to run it manually again.
Record every submission, quota check, and account switch账号别名、Chrome配置名、任务ID、片段、入口、动作、结果、额度快照、恢复提示原文、下次检查时间、下一步。
Do not hard-code quota, refresh cycles, or per-use cost. Read the real-time values from the account page or generation configuration panel each time.

For the specific SOPs for Gemini and Flow, see browser-sops.md。

7. Generate multi-segment video

Segment 1:

九宫格分镜图1（整张） + 产品图 + 视频提示词1。

Segment 2:

First complete segment 1 video and extract the real last frame, then generate the segment 2 storyboard; generating in advance is prohibited.
When generating the segment 2 storyboard, you must reference both the real last frame of segment 1 and the segment 1 storyboard, keeping the same character, hairstyle, clothing, scene, lighting, and camera, and changing only the action and product state in this segment; after acceptance, immediately upload分镜图 2。
Required: the real finished last frame of segment 1.
- Then add: the entire nine-grid image for this segment, the product image, and this segment's prompt.
At the beginning of the prompt, declare the image roles and priority: the last frame is the exact starting point, the storyboard image is the action/target reference, and the product image is only responsible for identity locking.
If Flow Frames mode can meet the requirements and there is a separately generated exact target last frame, use Start=上一段尾帧、End=精确目标尾帧to obtain stronger start/end control; do not cram the entire nine-grid image into End.
If you must reference the last frame, storyboard, and independent product image at the same time, use Flow Ingredients or Gemini Omni multi-image input.

8. Download, post-process, extract the last frame, and send it back

Download the local file immediately after each segment is completed.
Use ffprobe Verify that the file is readable, that its duration matches the actual single-segment duration displayed at the professional entry point at submission time, and that it is 9:16 portrait; the actual width must be less than the height. Landscape, square, rotated, or durations clearly inconsistent with the interface configuration are all judged as failures and must be redone.
When the Gemini output has a visible star-shaped watermark in the lower-right corner, keep the original file and preferentially call the installedgemini-watermark-remover Skill; if missing, first obtain permission via the environment precheck and install it. It is forbidden to use delogo, Gaussian blur, mosaic, overlay occlusion, or edge cropping to pass off as watermark removal.
- First check thegemini-watermark-remover Skill; if missing, follow the postprocess-sop.md instructions to state the source and install after obtaining the user's consent.
- Process batch videos one by one to avoid multiple local watermark-removal browser tasks competing for ports.
Extract the beginning, middle, and end frames from the watermark-removed file, and zoom in on the lower-right corner to confirm: the star-shaped mark has disappeared, the background texture and edges are continuous, and there are no rectangular blur blocks, residual sharp corners, or flickering.
Check the final video against the text post-production checklist from the breakdown stage. If core text is missing, has typos, garbled characters, wrong position, wrong appearance interval, or is occluded by the visuals, the text must be added or corrected in post-production; a final video that has not completed the text hook cannot be uploaded as a pending-acceptance finished product.
When adding text, preserve the original text, capitalization, spelling, punctuation, line breaks, position, style, and duration; do not polish on your own. Avoid duplicate overlapping text: text that already exists correctly in the final video must not be covered again. For specific methods and acceptance, see postprocess-sop.md。
Use ffprobe to confirm again that the post-processed version is still 9:16, that its duration is basically consistent with the original video, and that an audio track exists. For the complete command, tool boundaries, and acceptance rules, see postprocess-sop.md。
After completing segment 1 of a two-segment task, use ffmpeg to extract the last stable frame from the final post-processed file; do not fabricate the tail frame from the storyboard image. Immediately upload尾帧 1and read it back, as the basis for segment 2 continuation and cross-device continuation.
Handle according to the delivery method confirmed at startup: when immediate return is allowed, upload the final qualified post-processed file to the corresponding成品视频 ifield and read it back; when unified confirmation is required, save it only in the finished product directory and do not replace the Feishu attachment in advance.
If there is another segment, use the just-extracted real tail frame, the current storyboard image, and the product image to generate the next segment's storyboard; only after acceptance and upload, submit the next segment's video generation.
Before returning, first read the target finished product field to avoid duplicate appending on retry; upload using a relative path from the directory where the file is located.
After uploading, read back and verifyfile_token、nameandsize, then write最后成功步骤=成品视频 i 已上传；运行日志to record whether the text hook exists, whether text was added, and the acceptance result.
After all planned segments have been generated and passed machine acceptance, change the task to待验收, and do not directly change it to已完成。
Only after the user explicitly confirms the entire task's finished product, replace the old attachment the user rejected and change the status to已完成; then finally read back the planned segment count, all finished product attachments, and the status.
On failure, write失败原因, increase重试次数and change to失败; do not report partial success as complete.

Scheduled polling

By default, do not create a scheduled task.

When the user wants to continuously process new submission records and polling has not yet been configured:

First ask the user whether to set up scheduled reading of the Feishu multidimensional table.
After the user agrees, then ask about the polling interval and runnable time window, for example every hour, every 2 hours, or a specified time each day.
Use the automation tool provided by the product to create or update automated tasks; automated tasks only process records whose 'status is empty or pending and attachments are complete'.
Do not judge new tasks by creation date alone; status is the primary lock, attachment completeness is the admission condition, and record ID is used for deduplication and tracking.
At each scheduled run, first read the quota log: when Gemini is available, prioritize Gemini; when Gemini is restricted, switch to Flow; when Flow is also exhausted, recheck Gemini, and only wait for recovery after confirming it is still restricted.
When there are multiple accounts, follow account-routing.md read back the account table, check expired accounts, first poll all Gemini accounts, then poll Flow accounts; when all are unavailable, schedule only the earliest subsequent check.
On each run, append to运行日志the time, account alias, Chrome profile name, record ID, fragment, entry, action, result, quota snapshot, original recovery prompt text, and next step; do not merely overwrite the previous failure information.
If there is no automation tool, explain the limitation and provide an executable alternative; do not write a resident background script on your own.

Continue running after Gemini restriction recovery

When a Gemini usage restriction occurs:

Open设置 → 使用限制。
Read the current usage, short-cycle recovery time, and weekly quota recovery time displayed on the page; do not assume a fixed 3 items, 5 hours, or 24 hours. The user-observed 'about 3 items per session, about 5 hours' can only be used as an empirical note and cannot be used as a scheduling basis.
Only when the page gives a complete date, time, and time zone, write the converted accurate time into下次运行时间. If it only shows a time or明天再来and other information lacking a date/time zone, write the original text into最后成功步骤，下次运行时间leave it blank to avoid incorrect scheduling.
Change the status to等待额度恢复, and in最后成功步骤retain the current fragment, entry, usage, and original recovery prompt text.
Only when the recovery time can be parsed unambiguously and the user has already authorized automatic continuation, use the automation tool to schedule the next run 2–5 minutes after the recovery time, and carry the Base link, unfinished record ID, and fragment number.
If the user has not authorized it, or the recovery time lacks a date/time zone, tell the user the original page text and ask whether to schedule continuation.
When Flow is available, first continue consuming Flow's independent credits; do not let the task wait idly.

Feishu notification

By default, only update the Base status; do not proactively send messages.
When the user requests notification, uselark-imparse and confirm the individual or group chat target; do not guessopen_idorchat_id。
May be in待用户操作、等待额度恢复、待验收、失败and已完成send a brief notification including the product name, task ID, current fragment, status, next step, and Base link.
A successful Base write does not equal a successful message send; read back and report the two separately.

Completion check

Each planned fragment in the Base has a corresponding prompt, storyboard image, and finished video.
The storyboard image was sent back immediately after generation, not merely saved locally.
Segment 2 used the real last frame of segment 1.
The video is 9:16 and can be read and downloaded.
The Gemini visible watermark has truly been removed, there is no blurred block, and the original audio track still exists.
If the original work relies on on-screen text to achieve its hook, the final product already contains the exact text verified through frame sampling at the first frame, middle section, and end section; missing characters, wrong characters, or garbled text are all considered incomplete.
Gemini/Flow's actual entry point, model, duration, unit price, and recovery time come from the current interface, not from old memory.
The quota log has recorded the current entry point, two types of quota snapshots, the check time, and the next step.
For a single account, the actual account alias used has been recorded; for multiple accounts, the account selection, the reason for switching, each account's independent quota status, and the next check time have been recorded.
After machine generation ends, first enter待验收; only after user confirmation enter已完成。
Finally, re-read the records and report the number of completed segments, the entry point used, failure/retry situations, and tasks still pending.