Make every deconstruction answer three questions: which type of people's problem this video hits; how it keeps viewers watching step by step; and which methods are suitable for your own account.
This Skill is executed in four steps: input video, video parsing, shot breakdown, and full transcript. For sources, page-by-page fields, and parts not shown, see Original Document Verification; read it on first execution to avoid treating local supplements as the original framework.
Triggers and Division of Work
| Request | Scope of Execution |
|---|---|
| Deconstruct a specified peer video, viral content, or break down following the process | Run the four steps below and provide transferable methods |
| Review your own video | Also run the four steps, using the conclusions to pinpoint problems in specific segments |
| Only a transcript, screenshots, PDF, or partial materials | Complete the judgeable parts with the available evidence; state the specific gaps caused by missing video, comments, or timing information |
| Search accounts, search keywords, batch-scrape works | shengjiang-research is responsible for collection; this Skill is responsible for in-depth deconstruction of a selected work and does not automatically expand the account or platform scope |
| Optimize only the opening, analyze only communication psychology, revise only the script | Use according to the host registry dbs-hook / dbs-resonate / dbs-spread or the corresponding writing Skill; if these optional Skills are not installed, analyze directly within the user's limited scope without forcing the full process |
1. Input Video: Establish Evidence First
Obtain the link, original footage, subtitles, data, comments, and purpose from the user message and the current project. Reuse existing information directly; ask follow-up questions only when the work object or key objective cannot be determined. If the user has not stated their own positioning, complete the deconstruction of the original video first, and write only applicable conditions in the transfer section.
- Re-read the actual materials and record the platform, work title, author, original link, publish time, duration, collection time, and local original file location. Keep "not provided" for unknown fields; do not infer from file names.
- First confirm whether this is peer learning or self-review. Self-review needs to distinguish between "the people the user originally intended to reach" and "the people inferred from the existing content."
- Use the platform's original fields for public data: record likes, recommendations, saves, comments, shares/reposts, and views separately; "recommendations" must not be rewritten as views.
- "Viral hit" is a performance description pending verification. When only single-item data is available, present the actual performance; when relative performance needs to be judged, use existing comparable data from the same account and platform, and do not collect additional data without authorization. Without views, retention, or backend evidence, do not calculate view rate or completion rate, and do not present segment-by-segment speculation as measured drop-off.
- When only the work link is available, find and fully read from the Skills installed on the host
shengjiang-research(locate via the registry if one exists) and the true source it redirects to, and reuse its video download, subtitle/third-party ASR, details, and comment processes. If not installed, first follow the installation instructions Install; existing materials can continue to be analyzed. First display the specific request and fees, and execute according to existing authorization; the writing or testing of this Skill does not automatically authorize paid collection. - By default, only the specified work is processed, and comments follow the scope of the collection Skill. When the platform API is blocked, record the specific gap and continue analyzing existing materials; do not switch to web scraping to fill the gap.
- If reliable official subtitles already exist, reuse them directly; otherwise use the existing third-party ASR. Local speech models are prohibited. Titles and descriptions are not treated as verbatim transcripts.
- Prepare the complete video, subtitles with timestamps, and necessary visual evidence before analysis. Technical execution may obtain subtitles first; delivery is still organized according to the original four steps.
When a PDF or long screenshot contains materials, both the body text and the images must be read. Render the whole page to verify context, zoom in on embedded images to read fields, subtitles, and numbers; merge and understand screenshots spanning pages. Clearly record parts that are unclear or cropped out of the original image, and do not fill them in as original text.
2. Video analysis
Work performance and comprehensive diagnosis
First use one paragraph to explain how this video establishes viewing value, specifically clarifying the effective mechanisms and where the evidence is insufficient. Then fill in the six fields visible in the original image:
| Field | Question to answer |
|---|---|
| Content type | The content form and the task it undertakes |
| Core proposition | The main judgment the creator wants the audience to believe, distinct from the topic name |
| Target situation | Which type of people are in what situation, and what problems or emotions they have |
| Viewing promise | What answer, understanding, or action value can be obtained after watching |
| Core mechanism | How the opening, questions, viewpoints, evidence, and ending work together |
| Evidence boundary | Which are original video facts, author claims, comment feedback, or analytical inferences; what specifically is missing |
Each key conclusion points back to a timestamp, original sentence, image, or comment record. Data cited by the author only proves that he used that evidence; unverified data cannot automatically be treated as fact.
How the opening establishes a reason to watch
Determine the actual opening interval, and analyze it in combination with the first spoken segment, cover/on-screen title, and visible subtitles:
- Original video time and original sentence.
- Hook type and specific function; avoid merely applying labels such as "contrast, suspense, pain point".
- The information gap created: what the audience specifically wants to know at this moment.
- The promised viewing benefit.
- Fulfillment status: which later timestamp responds to it; the basis for full fulfillment, partial fulfillment, or non-fulfillment.
How the content is established step by step
Divide paragraphs according to semantic progression, not mechanically by fixed equal seconds. Record each paragraph: time range, core content/original sentence, expression method, changes in audience state (analytical inference), content task of this paragraph, and how it connects to the preceding and following text.
Based on the paragraph evidence, answer the questions raised in the original text: whom it targets, why they stay, what information gap is created, how the viewpoint is established, what is used to build trust, and what the audience agrees with and questions. The original screenshot does not display the complete names of the sub-structure columns, so do not fabricate the seven algorithms that are not shown.
Specific checks: whether the questions are gradually answered; whether conclusions are thrown out without intermediate explanation; whether examples and data support the conclusions; which sentence adds new information; which paragraph is repetitive, off-topic, or has an interrupted connection. When there is no real retention data, only say "reasons that may affect continued viewing".
What do the top comments reflect
First state the number of comments, collection time, sorting method, and scope. Sort by likes only when like counts are available; when this cannot be confirmed, say "comments provided". Keep representative original comment text, like counts (if any), and source IDs/locations, so that only a secondary summary does not remain.
Use the four fields shown in the image:What it confirms, doubts and resistance, user needs, emotional resonance points. One comment may support multiple observations; do not fabricate comments to fill sections. Then derive the next topic from unresolved questions that have evidence, stating the source comments and the facts that need to be supplemented.
Comment agreement only proves that someone expressed approval; the comment sample does not represent all viewers. Preserve the user's doubts in their original meaning; do not turn them into praise. If there are no comments, mark it as missing; do not use your own imagination as user feedback. When the PDF contains only a comment analysis summary, explicitly cite it as "secondary analysis in the material".
3. Storyboard: return to the actual footage
Segment by actual shot changes, subtitle or demonstration changes, and content task changes. Content segments and shot segments may differ; do not treat semantic segmentation as actual edit points.
Each segment includes start and end times, synchronized spoken narration, visual design, visible subtitles, content task, with keyframes or a rewatchable reference to the original footage. In the visual description, specify the shot size, positions of people/objects, demonstration actions, how text appears, and how the visuals help understanding.
When the original footage is available, first observe the complete timeline, and verify against keyframes and continuous clips near transitions; sparse frame sampling only supports the moments actually observed. For frame extraction, reuse the currently available FFmpeg/video tools, store temporary frames in the system temporary directory, and keep the small number of evidence images that need long-term citation in the project material directory.
When only static images are available, do not infer camera movement, gesture changes, editing frequency, or the continuous visuals across the whole segment; when only a script is available, do not generate a fabricated storyboard. Creative suggestions may be given separately, but they must be clearly labeled as suggestions.
Self-review check: when does the opening get to the question; do the visuals help understanding; where does it drag and why; do the subtitles highlight the key points; do the spoken narration, visuals, and subtitles serve the same task.
4. Full transcript: complete retention and rewatchability
Output a complete verbatim transcript with timestamps from official subtitles or third-party ASR results, preserving colloquialisms, repetitions, strong judgments, and questions; do not rewrite it into a condensed version without authorization. Note the source; separately flag suspected errors in names/brands/terminology, and keep a record of revisions after verification.
Check the actual subtitle coverage against the media duration, and look for missing beginning, middle, and ending. Silence or end-card visuals do not equal missing words; judge based on the original footage. Keep start and end times when available; when only a start time exists, do not fabricate an end time. Do not evenly distribute time over plain text to fake subtitles.
When the PDF screenshot shows only part of the verbatim transcript, the deliverable name must be "verbatim transcript excerpt"; other sections must not describe it as the full video. When a reliable transcription cannot be obtained, deliver the existing analysis and the specific gaps.
Transfer to your own content
This is a local execution supplement, serving the original text's requirement to "judge which methods suit you".
- Peer learning: list the specific mechanisms worth borrowing, evidence from the original footage, applicable conditions, how to use them in the target account, and the real materials needed. Preserve the tactical details and emotional function; do not copy the other party's identity, experience, income, clients, or product results.
- Self-review: check the target audience, the reasons to keep watching, the questions raised, the evidence for viewpoints, and the on-screen text/subtitles/voiceover item by item; output "specific passage → impact → next actionable adjustment". Explicitly retain the parts that have no problems; do not force problems just to fill a quota.
- The positioning and columns of the current IP are read from the project's source of facts and are not hard-coded as business state in this Skill. The user only wants a breakdown; do not rewrite the original draft, mark items as adopted, change the shooting schedule, or publish content without authorization.
- Do not set weights, total scores, viral probability, fixed formulas, or guaranteed-viral promises that do not exist in the original text. When a migration example is needed, use only verified self-owned materials; if materials are insufficient, provide structural and material requirements.
Delivery and Acceptance
Use the report template to organize the results of the four steps. By default, produce one complete Markdown report, citing existing original videos, subtitles, and data via links; if the user specifies HTML, tables, etc., produce them according to the corresponding Skill. Length varies with the material; fill in fields only when there is evidence, and explain gaps in a consolidated section.
Before writing, determine the single target according to the host's placement rules. Breakdowns of the current IP go in the corresponding benchmark material directory; other projects follow the project; this Skill directory holds only execution rules, references, and tests, and does not accumulate daily business reports.
Verify before completion:
- The object, data fields, source, and time are correct; facts and inferences are distinguishable.
- Check the body text, images/original video, comments, and subtitles according to the actually available scope; key conclusions have locators.
- The opening promise and its fulfillment, the paragraph-by-paragraph progression, the four comment items, and the five storyboard items are all addressed; missing materials are clearly noted.
- The storyboard and subtitle timings are compatible; the report does not treat screenshot analysis as having fully watched the original video.
- The original methods and local execution supplements are traceable; the migration does not misappropriate others' facts.
- The report, cited originals, and necessary evidence can be opened; update the most recent index and log as required by the host.
Completion status is divided into "complete breakdown" and "partial breakdown based on existing materials". Completing the Skill writing or offline testing does not mean that new video collection and the full process have passed. Test cases and real records are in evaluation cases and acceptance records。