What does Short Video Account Analysis do?
Complete an account analysis in one run as four local file groups, covering six tasks. The default sample is the 50 most recent posts sorted by actual publish date, and the default language is Chinese; if the user explicitly specifies a quantity, date range, or output location, follow the user's instructions.
Delivery Convention
Create the directory AccountAnalysis/<Account Nickname>_<Collection Date Time>/ on the desktop to avoid overwriting previous results.
01_Post Data and Rankings/: 50 raw posts with title, publication text, publication time, likes, link, author, and theme; sorted by likes; top 5 list. Provided in Markdown and CSV.02_Theme Analysis.md: Theme definitions, post count, percentage, average likes, and number of valid like samples; add median and extreme value effects when necessary.03_Top5 Transcripts/: Five separate Markdown files, filenames using the original post titles, containing the original title, original link, publication time, like count, collection time, author, transcription source and status, followed by the verbatim transcript.04_Persona Audience Product Analysis.md: Account bio, personal IP persona, target audience, product, and product selling points; each item supported by evidence, inference boundaries, and information gaps.
Also save original evidence and machine-readable state in source/ to allow resumption from breakpoints. These are the basis for the four groups above, not additional analysis tasks. See references/data-contract.md for detailed data structures and document formats.
Main Window and Luna Division of Labor
- The main window manages collection scope, task queue, evidence, exceptions, and final review. The user prefers Astra for oversight; only operate when the environment actually supports switching, and do not claim to have switched the current model without actual capability.
- Ordinary page collection, text entry, initial theme classification, and organization:
gpt-6-luna,xhigh. - Noisy audio, overlapping speakers, jargon-heavy content, subtitle-audio conflicts, omissions, or verification of digits and proper nouns:
gpt-6-luna,max. - When dispatching, set the actual model parameters of the tool: use
modelandreasoning_effortfor sub-agents, andmodelandthinkingfor standalone chats; simply writing the model name in the prompt does not constitute proper configuration. If tool limitations prevent full-history model coverage for all branches, use a context mode that allows overriding and explicitly pass the necessary materials. - Prioritize sub-agents. Each of the top 5 posts is handled by an independent task, queued according to actually available slots; each task uses its own directory. Only use the create-chat tool when the user explicitly requests a standalone chat, and follow that tool's authorization rules.
- If a specified model is unavailable, state this immediately; do not silently switch to another model or fabricate a successful setup.
- Luna is responsible for extraction, organization, and proofreading; audio recognition is performed by the ASR engine. Increasing the reasoning tier will not resolve download failures or single-queue ASR service limits.
Execution
1. Establish Evidence and Sample
Read the collection and classification sections of references/workflow.md. Parse the shared link to extract the real account, confirm the nickname, homepage link, bio, and collection time. Prioritize reusing verifiable collection materials from this run; old data of 100 posts must be re-filtered to the most recent 50 based on date, and old top 5 cannot be reused.
Use currently available official APIs, connectors, or browser capabilities. If the user specifies a browser like ego, first verify the tool can connect; only use data returned legally from visible page content or APIs. If encountering login walls, CAPTCHAs, inaccessible content, or incomplete samples, report the specific gap promptly and continue with unaffected parts.
Deduplicate post IDs. Keep pinned posts and sort them by actual publish date. Store title and publication text in separate fields; neither is equivalent to a spoken-word transcript. If publication date, exact likes, title, or bio are unattainable, retain empty values and do not fabricate.
2. Classification, Statistics, Ranking
Read references/data-contract.md, normalize the data, and assign one primary theme to each post. Establish a classification based on the account's content; do not assume a fixed number of categories. Low-evidence entries can be marked as "unclassified." First determine the set of 50 most recent posts, then rank within that set.
Use this skill's deterministic script to generate statistics and lists:
python3 /path/to/account-analysis/scripts/analyze_posts.py \
--input /path/to/collected_posts.json \
--output-dir /path/to/run --limit 50
Replace the example paths with actual absolute paths. After checking analysis_status.json, distribute 01_Post Data and Rankings/top5.json. Unknown likes must not be treated as 0; abbreviated likes can only be marked as approximate values, and if this might affect the ranking, label the top 5 as tentative. Do not produce statistical results by mental calculation.
3. Top 5 Transcription
Read references/transcription.md. Prioritize obtaining available subtitles. When a spoken-word transcript is required, compare against audio; existing subtitles do not guarantee full coverage of the spoken content. If no subtitles are available, obtain the audio and perform ASR. Use OCR to supplement visual information or handle segments lacking audio.
If the top 5 includes an article or image-text post, save the original text or arrange the on-screen text per page, noting the post type and source; if there is no spoken content, state this honestly. These posts retain their original rank; do not replace them with video posts just to have five videos.
For Douyin, the local douyin-video skill can be used; first read its documentation. Track media acquisition, ASR, and Luna proofreading separately. This skill's scripts/transcribe_local.py provides local service recognition, a serial queue, accurate job ID tracking, error logging, and resume caching; it does not turn a single-queue service into a concurrent service.
One collector operates the shared browser; one submitter manages the shared ASR queue. Verification and analysis of different videos with already-obtained text can be parallelized. Do not open multiple windows to compete for the same tab or re-submit the same video.
Output each post independently, preserving original wording, verbal repetitions, and uncertain sections. ASR completion does not mean the transcript has been proofread. If obtaining a transcript for any top 5 post fails, keep the original position and status; do not silently substitute the 6th post. If the transcript does not cover the entire video, explicitly write "partial transcript" and do not claim completeness.
4. Account Analysis and Final Review
Read the evidence and reporting sections of references/workflow.md. Combine the homepage, 50 post titles and texts, top 5 transcripts, and accessible product pages to complete four analysis chapters. If product or actual user data is insufficient, retain the chapter and list unknown items and missing evidence.
The main window checks:
- The time range of the 50 most recent posts, pinned post deduplication, collection coverage status, and whether the like collection time is clearly stated.
- Whether the total count of primary themes equals the sample size, whether percentages are calculated against the full sample, and whether unknown likes are excluded from the mean and disclosed.
- Whether the top 5 matches the ranking; whether all five files contain the original title and five metadata items, and whether the original link corresponds to the post ID; whether transcription status matches the coverage evidence.
- Whether self-descriptions, observable facts, analytical inferences, and unknowns are clearly separated; do not guess actual age, income, product sales, or conversion rates from content preferences.
- Whether all four groups are complete, and whether all gaps were reported when discovered and finally listed together.
Finally, provide the desktop directory and links to key files, briefly summarizing the number of themes, the completion status of top 5 transcription, and analysis limitations. If partially failed, deliver completed results and specific entry points for continuation; do not mark the entire task as complete.