tiktok-research（抖音研究）

TikTok Research Skill

Use this skill for TikTok metadata research:

• keyword and hashtag sampling
• competitor or creator sampling
• creator account discovery
• profile enrichment
• comments and audience language
• trend and benchmark discovery
• dataset summaries

Follow shared routing and guidance rules in:

• postplus-shared release-shell rules
• postplus-shared research preferences

Read these references before implementation:

• ${CLAUDE_SKILL_DIR}/references/actor-selection.md
• ${CLAUDE_SKILL_DIR}/references/normalized-schema.md
• ${CLAUDE_SKILL_DIR}/references/tool-contracts.md

Do not treat generic "TikTok research" requests as plain web search unless the user clearly wants external articles, policy, or news.

Cost Discipline

When using free-tier collection credits, prefer:

• serial runs, not parallel runs
• 5-12 results per hashtag or query
• 3-5 seed hashtags per batch
• 3-5 strongest search queries per batch
• comments only after a promising batch is identified

Do not brute-force full-market collection first. Sample, inspect, refine, then expand.

Current bounded-first-pass rule:

• treat brief limit as the target shortlist or candidate goal, not as the per-query scrape size
• let build_tiktok_actor_input.mjs keep first-pass resultsPerPage / maxProfilesPerQuery bounded unless you explicitly need a broader second pass
• for tiktok-scraper, default to searchSection: "/video" for content-first creator discovery unless you intentionally want account-search behavior
• keep comment collection, media downloads, subtitles, follower/following expansion, and similar extras off in the first pass unless the current step truly needs them

Key Files

• ${CLAUDE_SKILL_DIR}/scripts/collection_actor_run.mjs
• ${CLAUDE_SKILL_DIR}/scripts/build_tiktok_actor_input.mjs
• ${CLAUDE_SKILL_DIR}/scripts/normalize_tiktok_dataset.mjs
• ${CLAUDE_SKILL_DIR}/scripts/rank_tiktok_accounts.mjs
• ${CLAUDE_SKILL_DIR}/scripts/expand_tiktok_creator_graph.mjs
• ${CLAUDE_SKILL_DIR}/scripts/analyze_tiktok_dataset.mjs
• ${CLAUDE_SKILL_DIR}/scripts/collect_top_video_comments.mjs
• ${CLAUDE_SKILL_DIR}/scripts/analyze_tiktok_comments.mjs
• ${CLAUDE_SKILL_DIR}/scripts/generate_tiktok_report.mjs
• ${CLAUDE_SKILL_DIR}/templates/ — sample hashtag, search, and competitor input files

Prerequisites

In the product-shell runtime, follow postplus-shared release-shell rules.

TikTok-specific runtime notes:

• when this skill references its own scripts, references, or templates, use absolute paths anchored at ${CLAUDE_SKILL_DIR}; do not rely on bare scripts/..., references/..., templates/..., or skills/20-research/tiktok-research/... paths that depend on the current working directory
• when extracting creator handles from a normalized video dataset, use .items[].authorUsername
• if hosted capability is unavailable, unauthorized, or returns a stable network error, stop immediately instead of switching to ad hoc shell glue

Node 18+ is required. This workspace already has Node available.

Default Actors

• creator discovery by matched videos: tiktok-scraper
• keyword account-search supplement: tiktok-user-search-scraper
• profile enrichment: tiktok-profile-scraper
• keyword / hashtag video discovery: tiktok-scraper
• direct video enrichment: tiktok-video-scraper
• hashtag-only sampling: tiktok-hashtag-scraper
• comment collection default: tiktok-comments-scraper
• low-cost search expansion fallback: tiktok-scraper
• low-cost exploratory query fallback: tiktok-scraper-api
• TikTok Shop creator discovery: tiktok-shop-creators

Treat these as defaults, not hard requirements.

Release-shell note:

• tiktok-profile-scraper (secondary provider) remains shelved for current product-shell runs. Do not route to it until billing adds and verifies a provider cost rule.

Routing Rules

Start from the user's real collection surface, not from one favorite actor.

• creator search with follower-band or content-fit constraints -> tiktok-scraper first
• creator search as account-search supplement -> tiktok-user-search-scraper
• profile enrichment -> tiktok-profile-scraper
• bulk profile enrichment stays on tiktok-profile-scraper
• topic/search video discovery -> tiktok-scraper
• hashtag-only sampling -> tiktok-hashtag-scraper
• known video URL enrichment -> tiktok-video-scraper
• focused comments -> tiktok-comments-scraper
• larger or cheaper comment runs -> tiktok-comments-scraper
• region-aware low-cost discovery -> tiktok-scraper
• shop creator analytics -> tiktok-shop-creators

Do not use TikTok Shop actors for generic creator discovery. Do not use Creative Center ad actors as if they were organic creator data. If the user wants paid ad intelligence, route to tiktok-ad-research. If the user specifically wants TikTok music or sound discovery, route to:

• trending music by region -> use the music-discovery path within tiktok-research
• videos under a specific music/sound -> tiktok-music-sound-collector when released
• local download and audio extraction -> tiktok-music-archive-downloader

For multi-step TikTok music workflows, read postplus-shared TikTok music workflow.

Creator Discovery V2 Rule

For TikTok creator discovery, do not default to plain user search when the brief includes:

• follower bands like 5k-10k
• recent activity
• local language or geo fit
• audience fit inferred from what creators actually post

Prefer this order:

1. collect matched videos
2. extract authors
3. expand around strong seed videos if recall is weak
4. enrich profiles
5. rank with profile plus content evidence

Use tiktok-user-search-scraper as a supplement, not the default first pass, unless the user explicitly wants account-search recall.

Input Compilation

When the request is expressed as a research goal instead of actor JSON:

1. write a small brief JSON with fields such as task, queries, hashtags, usernames, urls, limit, country
2. run ${CLAUDE_SKILL_DIR}/scripts/build_tiktok_actor_input.mjs
3. inspect the compiled actor input before the collection run if the request is high-cost or ambiguous

This keeps the workflow stable even when actor input shapes differ.

Bounded brief rule:

1. keep the first pass small by default:

- 3-5 strongest queries - 2-4 seed hashtags - limit set to the candidate or shortlist goal, not the per-query scrape size

1. only add explicit actor-size overrides such as resultsPerPage, maxProfilesPerQuery, commentsPerPost, or scrapeRelatedVideos when the current step truly needs them
2. if the first pass is too thin, expand in a second pass instead of starting broad

Product-shell temp-file flow

Inside the product-shell runtime, prefer this order:

1. when possible, skip the temporary brief file and call build_tiktok_actor_input.mjs directly with repeated flags such as:

- --query - --hashtag - --username - --url - --limit - --country - --actor - --output

1. write the actor input JSON into the current work folder's .postplus/ directory
2. keep raw datasets and other intermediate files in .postplus/
3. keep only final user-facing deliverables outside .postplus/

Direct-output rule:

1. ${CLAUDE_SKILL_DIR}/scripts/build_tiktok_actor_input.mjs --output <actor-input.json> creates that actor-input file itself
2. do not Read or Write the actor-input output path before running that script
3. do not pre-read the actor-input output path just because it does not exist yet
4. let the script create the file, then consume that same file path in the next step

Fallback only when the direct flag path is genuinely awkward:

1. write the brief JSON into the current work folder's .postplus/ directory with the Write tool
2. run build_tiktok_actor_input.mjs against that .postplus/*.json brief

First move for clear KOL discovery

For a clear request like:

• 帮我在 TikTok 上找 20 个适合卖护肤仪的 KOL，优先微中腰部账号，给我一个可继续筛选的 shortlist。

default to this first move:

1. skip the temporary brief JSON
2. run ${CLAUDE_SKILL_DIR}/scripts/build_tiktok_actor_input.mjs with --query, --limit, --actor, and --output
3. place the output under the current work folder's .postplus/
4. then run ${CLAUDE_SKILL_DIR}/scripts/collection_actor_run.mjs against that output file

Do not start by reading or writing a new .postplus/*.json file when the direct actor-input path already fits.

Profile-enrichment bridge

For creator discovery from a normalized videos dataset:

1. extract usernames from .items[].authorUsername
2. write a real profile-request JSON file under .postplus/
3. run collection_actor_run.mjs with tiktok-profile-scraper
4. normalize the profile dataset
5. run rank_tiktok_accounts.mjs --profiles <profiles-normalized.json> --videos <videos-normalized.json>

Do not call rank_tiktok_accounts.mjs with only --videos.

Merge rule

If you collect ranked results in multiple batches:

1. write each ranked batch to its own JSON file
2. merge them from disk with a plain file-based command such as:

jq -s 'map(.items) | add | sort_by(-.followersCount) | .[0:20]' file-a.json file-b.json

Minimal Workflow

Hashtag sampling

Build input:

node ${CLAUDE_SKILL_DIR}/scripts/build_tiktok_actor_input.mjs \
  --brief ${CLAUDE_SKILL_DIR}/templates/hashtags.json \
  --actor tiktok-scraper \
  --output <work-folder>/.postplus/hashtags-input.json

Run actor:

node ${CLAUDE_SKILL_DIR}/scripts/collection_actor_run.mjs \
  --collection-key tiktok-videos \
  --input <work-folder>/.postplus/hashtags-input.json \
  --output <work-folder>/.postplus/hashtags.json

Then summarize:

node ${CLAUDE_SKILL_DIR}/scripts/normalize_tiktok_dataset.mjs \
  --input <work-folder>/.postplus/hashtags.json \
  --actor tiktok-scraper \
  --dataset-type videos \
  --output <work-folder>/.postplus/hashtags-normalized.json

node ${CLAUDE_SKILL_DIR}/scripts/analyze_tiktok_dataset.mjs \
  --input <work-folder>/.postplus/hashtags-normalized.json

Search-query sampling

Build input:

node ${CLAUDE_SKILL_DIR}/scripts/build_tiktok_actor_input.mjs \
  --brief ${CLAUDE_SKILL_DIR}/templates/searches.json \
  --actor tiktok-scraper \
  --output <work-folder>/.postplus/searches-input.json

Run actor:

node ${CLAUDE_SKILL_DIR}/scripts/collection_actor_run.mjs \
  --collection-key tiktok-videos \
  --input <work-folder>/.postplus/searches-input.json \
  --output <work-folder>/.postplus/searches.json

Then summarize:

node ${CLAUDE_SKILL_DIR}/scripts/normalize_tiktok_dataset.mjs \
  --input <work-folder>/.postplus/searches.json \
  --actor tiktok-scraper \
  --dataset-type videos \
  --output <work-folder>/.postplus/searches-normalized.json

node ${CLAUDE_SKILL_DIR}/scripts/analyze_tiktok_dataset.mjs \
  --input <work-folder>/.postplus/searches-normalized.json

Competitor sampling

{
  "profiles": ["grammarly", "notionhq", "raycastapp"],
  "resultsPerPage": 10
}

node ${CLAUDE_SKILL_DIR}/scripts/collection_actor_run.mjs \
  --collection-key tiktok-profiles \
  --input <work-folder>/.postplus/competitors.json \
  --output <work-folder>/.postplus/competitors-results.json

Then normalize and rank:

node ${CLAUDE_SKILL_DIR}/scripts/normalize_tiktok_dataset.mjs \
  --input <work-folder>/.postplus/competitors-results.json \
  --actor tiktok-profile-scraper \
  --dataset-type profiles \
  --output <work-folder>/.postplus/competitors-profiles-normalized.json

node ${CLAUDE_SKILL_DIR}/scripts/rank_tiktok_accounts.mjs \
  --profiles <work-folder>/.postplus/competitors-profiles-normalized.json

Creator discovery by keyword

{
  "searchQueries": ["ai tools", "productivity workflow", "gmail tips"],
  "maxProfilesPerQuery": 8
}

node ${CLAUDE_SKILL_DIR}/scripts/collection_actor_run.mjs \
  --collection-key tiktok-users \
  --input <work-folder>/.postplus/creator-search.json \
  --output <work-folder>/.postplus/creator-search-results.json

node ${CLAUDE_SKILL_DIR}/scripts/normalize_tiktok_dataset.mjs \
  --input <work-folder>/.postplus/creator-search-results.json \
  --actor tiktok-user-search-scraper \
  --dataset-type user-search \
  --output <work-folder>/.postplus/creator-search-normalized.json

node ${CLAUDE_SKILL_DIR}/scripts/rank_tiktok_accounts.mjs \
  --profiles <work-folder>/.postplus/creator-search-normalized.json

Creator discovery by matched videos and graph expansion

Use this when the request cares about small creators, real topical activity, or content fit more than user-search ranking.

node ${CLAUDE_SKILL_DIR}/scripts/collection_actor_run.mjs \
  --collection-key tiktok-videos \
  --input <work-folder>/.postplus/topic-video-search.json \
  --output <work-folder>/.postplus/topic-video-search-raw.json

node ${CLAUDE_SKILL_DIR}/scripts/normalize_tiktok_dataset.mjs \
  --input <work-folder>/.postplus/topic-video-search-raw.json \
  --actor tiktok-scraper \
  --dataset-type videos \
  --output <work-folder>/.postplus/topic-video-search-normalized.json

node ${CLAUDE_SKILL_DIR}/scripts/expand_tiktok_creator_graph.mjs \
  --input <work-folder>/.postplus/topic-video-search-normalized.json \
  --output <work-folder>/.postplus/topic-graph-raw.json \
  --top 10 \
  --results-per-seed 6

Comment sampling

node ${CLAUDE_SKILL_DIR}/scripts/collect_top_video_comments.mjs \
  --input <work-folder>/.postplus/searches.json \
  --output <work-folder>/.postplus/comments.json \
  --collection-key tiktok-comments \
  --top 8 \
  --max-comments 40

Then summarize:

node ${CLAUDE_SKILL_DIR}/scripts/analyze_tiktok_comments.mjs \
  --input <work-folder>/.postplus/comments.json

Report output

node ${CLAUDE_SKILL_DIR}/scripts/generate_tiktok_report.mjs \
  --keyword-summary <work-folder>/.postplus/searches-summary.json \
  --hashtag-summary <work-folder>/.postplus/hashtags-summary.json \
  --competitor-summary <work-folder>/.postplus/competitors-summary.json \
  --comments-summary <work-folder>/.postplus/comments-summary.json \
  --output reports/tiktok-research.md

Default Sequence

1. choose the route first: content-first, handle-first, graph-first, or mixed
2. collect 3-5 seed hashtags or 5-10 jobs-to-be-done searches
3. extract authors from matched videos before relying on account search
4. expand around strong seed videos when recall is weak
5. enrich a shortlisted set of profiles
6. rank with both profile and video evidence
7. summarize each dataset locally
8. merge findings into hooks, formats, and content pillars

Skill Handoff

Escalate to skills/40-creative/video-analysis when:

• the user wants hook or structure breakdowns
• the user asks why specific videos work
• the user wants shot-level or spoken-line analysis
• the user wants to recreate or adapt benchmark videos

When the request is broad, ask one short guiding question before running:

• "你是想先看 TikTok 上哪些方向值得研究，还是已经有几条视频想让我直接拆内容结构？"

Mention that this skill finds good samples first, and video-analysis can read the actual videos later.

Escalate to skills/20-research/tiktok-ad-research when:

• the user wants Creative Center top ads
• the user wants paid ad hooks, objectives, or ad brief inputs
• the request is about ad-market scanning rather than organic creator or content behavior

Download Recovery Note

This skill is primarily for research, not guaranteed video asset persistence.

If a later workflow needs the actual benchmark videos and the earlier local files are gone:

1. recover from saved local video paths if available
2. otherwise download directly from the TikTok post URLs with yt-dlp
3. store the recovered videos in a stable workspace folder
4. hand those files to video-analysis

Do not assume collection dataset items will always contain reusable mediaUrls or downloadAddr fields.

Good Output

For each run, extract:

• strongest opening lines
• repeated framing patterns
• recurring hashtags
• recurring creators
• bio-link or outbound-link signal on creator profiles
• comments that reveal pain points or objections
• source surface and source query so later analysis can distinguish search vs hashtag vs profile vs direct URL runs
• region or location signal when the actor supports it
• creator-discovery evidence:

- which queries surfaced each creator - how many matched videos each creator contributed - whether the creator was found directly or through graph expansion - whether the profile has contactable public signals

• visual format types:

- screen demo - before/after - POV/relatable - listicle - talking head