9398-haha/douyin-benchmark

Douyin Benchmark

Let the script handle public account data collection, aggregation, and HTML rendering; the agent writes a Markdown report with evidence boundaries based on report_data.json.

Boundaries

• Only collect publicly visible web data; do not crack signatures, bypass CAPTCHAs, or circumvent access controls.
• Only use the ~/.douyin-benchmark/chrome-profile dedicated login state. Do not read, copy, or directly point to the user's daily Chrome user directory; do not output cookie values.
• Keep missing values as missing; do not rewrite them as 0. Likes can only serve as a proxy for public engagement, not as a measure of play performance, interaction rate, or follower conversion.
• Distinguish between a regular login form, real security verification, and page service errors. Stop on encountering a security challenge, page service error, or empty response soft block; do not retry in a loop, auto-switch IPs, create delayed tasks, or promise to complete later.

Initial Setup

Run the following commands in this Skill's directory:

npm ci
node scripts/login.js

login.js opens a dedicated browser, authenticates to a real account entry point, then automatically saves and closes. To check an existing login state, run node scripts/login.js --check.

One-Command Execution

Prefer to ask the user for the Douyin App "Share Profile" short link. A numeric Douyin ID is not the same as sec_uid; directly concatenating /user/<number> may silently redirect to the feed.

OUT="$HOME/douyin-benchmark-output/<account-identifier>"
node scripts/run_benchmark.js "<profile-share-link/sec_uid/douyin-id/nickname>" "$OUT" \
  --expected-id "<douyin-id, optional>"

By default, it uses headed Chrome for better observability and a real browser UA. Only explicitly add --headless in headless environments. The script executes in order:

1. Resolve the short link or sec_uid; only use in-app search when given a nickname or Douyin ID.
2. Collect profile info, posts, and collections, writing them to account_full_raw.json.
3. If posts exist, generate report_data.json and report.html.
4. Update run_status.json throughout; do not continue analysis after a failure.

Check Status First, Then Write Conclusions

Read run_status.json and account_full_raw.json.capture:

• complete: The API returned has_more=0. A full account-level report can be written; still compare the homepage post count with the collected count.
• partial: Only write a "sampled post report"; indicate the sample size, total homepage post count, and coverage near the title. Do not use terms like "full" or "account's max."
• blocked_empty_body: The posts endpoint returned HTTP 200 with an empty body. Stop requests, deliver only the profile snapshot and diagnostics; any retry must be user-initiated, and at most once.
• challenge: Ask the user to complete the verification in the dedicated browser before retrying; do not attempt to bypass.
• login_required: The page shows a regular login form, not a security challenge. Run the login script, then retry only once after success.
• page_service_error: The posts area explicitly shows "Service error, please refresh." Stop refreshing; after verifying the login state, retry at most once.
• no_post_response: May be an incorrect input or a page/API change. First check the sec_uid and the headed page; do not blindly increase scroll count.
• endpoint_error: The posts endpoint returned an HTTP/API error or non-JSON content. Keep diagnostics and stop analysis; first check for page/API changes.

Writing the Markdown Report

Only when report_data.json exists and contains analyzable posts, fully read the references/report-template.md and write $OUT/report.md following the nine-section structure.

Classify content into three types and use precise wording:

• Facts: Public fields from the JSON and their computed results.
• Observations: Titles, bios, collection names, cover images, or content of videos actually opened.
• Inferences: Persona, monetization path, content formula; write down evidence and confidence level, do not present as fact.

Do not write "popular comment sentiment" without comment text; do not claim camera work, voiceover, acting, or edit details without having watched the video. If the user specifies their own niche, then selectively transfer the replicable list.

Verification

• account_full_raw.json, run_status.json, report_data.json, report.html, and report.md must have consistent statuses.
• Both HTML and Markdown should display the actual collected sample size, total homepage post count, coverage, and capture_status.
• On an error, check capture.diagnostics.page: only record the title, path, login/challenge/service-error boolean flags, and video link count; do not save full page content or cookies.
• Numbers must match report_data.json item by item; missing values display —/Unknown.
• The HTML file must be openable offline; external text must be escaped. Failed cover images do not affect the main content.
• No horizontal overflow on the HTML page; the latest post titles must be readable. Posting frequency should show the last 12 months based on individual posts; only fill missing months with 0 if a full capture was performed.
• The Markdown report must have all nine sections, including data sources, evidence boundaries, a replicable list, and items that should not be blindly copied.

Configuration

• DOUYIN_PROFILE_DIR: Dedicated login state directory; default is ~/.douyin-benchmark/chrome-profile.
• DOUYIN_BROWSER_CHANNEL: Default is chrome; set to chromium to install Playwright Chromium first.
• DOUYIN_HEADLESS=true: Explicitly enable headless mode; default is headed.
• DOUYIN_SCROLL_WAIT_MS: Minimum scroll interval, default 1600ms; do not use it for high-frequency scraping.