git:20260903.e492851 to git:20260905.4909a33

97 added, 73 removed. Audit A to A.

---

name: trend-tiktok

title: "TikTok Trend Harvester Playbook"

- description: "Platform-specific intel for harvesting TikTok trends — Creative Center as the primary instrument (Trends tabs for hashtags/songs/creators/videos with country+industry filters, Keyword Insights, Symphony Assistant), the measured 10-day content lifespan (96% of reach) as the urgency anchor, sound-vs-format lifecycle asymmetry, scoring built on views-to-follower ratios (follower count is officially not a ranking factor), the fabricated-stat blacklist, and dead patterns from the For You eligibility standards. Activates in browser sessions whenever the user names TikTok."

+ description: "Harvest TikTok trend evidence into a source-linked brief with audience fit, observed momentum, original angles, and commercial-use constraints. Activate for TikTok trend scans, Creative Center research, and browser-based campaign discovery; separate observed posts from ad examples and generated suggestions."

license: Apache-2.0

- compatibility: "Octoweb browser access. Creative Center (creativecenter.tiktok.com) is public and free but heavily JS-rendered — needs a live browser session, not fetch; TikTok web works logged-out for search/tags/sounds."

+ compatibility: "Requires browser and network access; authenticated TikTok business or publishing account access when a surface requires it."

capabilities: octoweb memory-read memory-write

domains: browser

rules:

- session(trend) content(tiktok)

- match(\btiktok\s+(trend|trends|harvest|brief)\b)

- match(\b(harvest|scan|analyze)\s+tiktok\b)

- session(trend) match(\bcreative\s+center\b)

---

## Overview

- This skill carries the TikTok-specific mechanics the trend-harvesting agent needs — harvest surfaces (Creative Center first), scoring signals, lifecycle timing, dead patterns, and a fabrication warning unique to this platform's data ecosystem. The agent owns the shared DNA loop (memory → harvest → score → cluster → DNA → hook bank → brief); this skill plugs in the TikTok parameters.

+ Collect inspectable TikTok evidence for a specific audience and buying situation. Return source-linked candidates with momentum, limitations and an original contribution; drafting and publishing are downstream work. Reuse prior observations only when their timestamp and filters are known.

## Mental model

- - Distribution is interest-cluster matching: the For You feed drives 72.7% of video views (Metricool 2026, n=2.31M posts) and TikTok states verbatim that follower count and past viral hits are "not direct factors" in recommendations. Consequence for harvesting: a micro account's breakout is pure algorithm signal — there is no follower-inertia to discount, unlike Instagram or X.

- - The urgency anchor is measured: 96% of a post's lifetime reach and ~98% of its interactions land within the first 10 days (Metricool). A "trending" post older than ~10 days is an archaeology find, not a live trend.

- - The feed deliberately diversifies (won't show two same-sound videos in a row) — saturation hits distribution mechanically, not just aesthetically.

- - Completion is the strong indicator per TikTok's only official weighting statement (finishing a longer video outweighs weak signals). High view counts on hyper-short loops are cheaper than the same counts on 60s+ videos — weight accordingly.

- - Sound-driven trends have a narrow participation window; format/meme-driven trends last materially longer (consistent practitioner observation, no measured study). Classify every trend as sound-locked or format-locked before recommending timing.

- - 2026 context: volume +72% YoY, views −31% — everything is more saturated than the same signal would have meant a year ago. Raise the bar.

+ Separate observed posts, commercial examples and generated suggestions. A popular post is evidence of attention; it does not establish retention, conversions or the cause of its distribution.

- ## Rules

+ TikTok's historical explainer says follower count and prior hits are not direct recommendation factors (official, TikTok recommendation explainer, 2020-06). That does not make views divided by followers a pure algorithm score. Account history and outside exposure remain possible explanations; don't infer current launch-pool rules from that old explainer.

- ### Harvest surfaces (Creative Center first, then in-platform)

+ Use chart changes and new relevant executions to assess momentum. A post-level reach window does not establish a trend's expiration date. Classify sound-dependent versus reusable-format candidates to decide what needs rechecking before production (directional).

- | Surface | URL | Yields |

- |---|---|---|

- | Creative Center — trending hashtags | `https://ads.tiktok.com/business/creativecenter/inspiration/popular/hashtag/pc/en` | Ranked hashtags, filter by country + industry + period |

- | Creative Center — trending songs | `https://ads.tiktok.com/business/creativecenter/inspiration/popular/music/pc/en` | Rising vs peaked sounds; "Approved for Business Use" filter |

- | Creative Center — trending videos | `https://ads.tiktok.com/business/creativecenter/inspiration/popular/pc/en` | What TikTok itself showcases as working now |

- | Creative Center — keyword insights | `https://ads.tiktok.com/business/creativecenter/keyword-insights/pc/en` | Search-demand signals for topic angles |

- | TikTok search | `https://www.tiktok.com/search?q=<term>` | Niche posts as ranked by TikTok search |

- | Tag page | `https://www.tiktok.com/tag/<tag>` | Post volume + top posts per hashtag |

- | Sound page | `https://www.tiktok.com/music/<slug>-<id>` | Videos using a sound, recency spread |

- | Anchor profile | `https://www.tiktok.com/@<handle>` | Niche anchors' recent output, view counts on grid |

+ ## Harvest procedure

- Set the country and industry filters FIRST in Creative Center — trends vary sharply by both, and unfiltered global charts are noise for niche work. Pages are heavily JS-rendered: navigate in the browser, wait for render, snapshot, scrape, close. Symphony Assistant (in Creative Center, logged-in) answers "show me top-performing <industry> content in <market>" in natural language — use it when charts are ambiguous.

+ ### Establish the scope

- ### Timing — use in-product state, not day-counts

+ Record target country, language, niche, buyer situation, account type, and whether the brief is organic, commercial or paid. Clarify what decision the harvest must support: topic selection, an opening, a demo format, a creator shortlist, or an objection-answer post.

- No measured trend-lifecycle study exists; circulating day ranges disagree (5–14 days peak; one claim of 72-hour peaks). Don't fabricate a window. The checkable rule: enter while the sound/hashtag is still RISING in Creative Center's trend charts and before it plateaus — chart trajectory is observable in-product on every harvest. Combine with the measured 10-day content lifespan: if the top posts on a sound are mostly >10 days old, its live phase is over regardless of chart rank. Sound-locked trends get "this week or skip"; format-locked trends tolerate later, better-crafted entries.

+ Read prior observations for matching filters when available. Don't write to persistent memory unless that workflow is authorized; return reusable observations in the brief otherwise.

- ### Scoring rubric (TikTok-specific signals)

+ ### Navigate from verified entry points

- Virality axis 0–5:

- - Views-to-follower ratio on the poster — the cleanest signal on any platform here, since followers officially don't drive distribution. A 5K-follower account at 2M views is a maximal structural signal. Tag account bands (micro <50K / mid / mega).

- - Comments-to-views — likes are the weak indicator; comment density marks the content that provoked. The measured comment asymmetry (questions +26% comments) means question-shaped winners are repeatable, not lucky.

- - Completion proxy — length × views: sustained views on 60s+ videos outrank the same views on 8s loops (completion is the official strong indicator).

- - Chart trajectory — rising in Creative Center beats absolute rank.

+ | Surface | Entry and use |

+ |---|---|

+ | TikTok One Creative Suite | Start at `https://ads.tiktok.com/creative/creativeCenter` and open the current Trends surface. Trends and Symphony Creative Studio are exposed here (official, Creative Suite, undated; verified 2026-09). |

+ | Trend charts | Follow live navigation for available topics, hashtags or music; record the actual URL, displayed filter labels and selected values. Don't assume every surface supports country, industry and period filters. |

+ | TikTok in-product discovery | Navigate search and open relevant results, then follow their tag, sound or profile links. Save the resolved URLs. Avoid constructed route templates or assuming logged-out access. |

+ | Keyword-related surfaces | Use a keyword tool only if present and its metric definition is visible. Record whether it describes paid creative, organic queries or something else; don't substitute ad keyword performance for organic search demand. |

+ | Content Suite | If authorized access is available, inspect existing brand-relevant UGC using AI Search, Projects and Creators (official, Content Suite announcement, 2026-06). Keep permission status separate from discoverability. |

+ | Symphony Creative Studio | Symphony Agent can generate briefs and ads from campaign prompts and trend/top-ad insights (official, Product Preview, 2026-07). Treat output as suggestions to verify against actual posts, never as observed examples. |

- Niche-fit axis 0–5 — same scale the agent applies on every platform; the algorithm matches interest clusters, so off-niche trend-chasing underperforms structurally (relevance outweighs follower count — Metricool's small-account growth finding). Drop everything below 3 on either axis.

+ Use the current root navigation when an old URL fails, redirects, or shows a different page. Record the final destination and access state. If navigation or labels have changed, describe the surface actually observed; don't claim a specific redirect history without observing it. If access is blocked, mark that coverage unavailable and continue with accessible evidence. Don't invent missing chart values.

- ### Trend taxonomy (maps to Creative Center's own tabs)

+ Render the page before extracting it. Record session market and login state without collecting credentials. Close only the research tabs you opened when finished.

- Sounds/songs · formats and memes · challenges · effects · hashtags. Classify every cluster; the class determines the window (sounds narrow, formats long) and the brief's original-layer requirement (see below).

+ The Commercial Music Library is described as supplying paid-ad-cleared tracks and an approved-for-business-use filter in the ad playbook (official, SMB Playbook, undated; verified 2026-09). Record the current rights/filter wording and selected track's permissions. Popularity is not commercial clearance. Do not assume the legacy label still appears on every chart.

- ### The originality gate

+ ### Capture source records

- TikTok's For You eligibility standards exclude "reproduced or unoriginal content… without any new or creative edits," and enforcement escalated 15 September 2025 (hits earnings AND visibility). Every recommended angle must name its original layer — the niche translation, take, or creative edit. Recommending "do this trending format" verbatim is recommending ineligible content. Watermarked cross-posts (including other editing apps' logos) are dead on arrival.

+ For each candidate, record:

- ### Data hygiene — this platform's special hazard

+ - Observation timestamp, resolved URL, handle, publication date or displayed age, and language/market context.

+ - Visible views and engagement counts, follower count when visible, duration and format. Mark unavailable data explicitly rather than estimating it.

+ - Exact hook transcription, first-frame description, overlay wording, caption, sound attribution and the visible disclosure state.

+ - Account's comparable recent posts; whether the candidate is a paid example, creator post, or ambiguous. Mark outside promotion unknown when unobservable.

+ - Relevant comment excerpts with enough context to explain the buyer's question; omit unnecessary personal information.

+ - Source surface, filters and metric definitions, including whether counts are rounded or collected at different times.

- A large fraction of 2026 "TikTok statistics" is AI-generated stat-farm content with fabricated figures under fake attributions to real firms (Sprout, Later, even TikTok itself). Known-fabricated families: hook-retention thresholds ("60% past 3s = 4x reach"), duet/stitch uplift percentages, per-video watch-time averages, AI reach penalties, numeric batch-test thresholds ("200-view jail"). Rule: figures enter a brief only from the platform's primary pages, the named measured studies (Metricool, Socialinsider, Buffer, Rival IQ/Quid), or the harvest itself. When citing a number in the brief, name its source or drop the number.

+ Label harvested numbers inline as (measured, live TikTok observation, n=observed sample scope, observation month); insert the actual scope and date. Keep exact timestamps in the record. Use (official, source name, publication or verification month) for documented mechanisms and (directional) for editorial interpretations without numerical claims.

- ### Dead patterns (don't recommend, discount when harvested)

+ ### Compare and cluster

- - Watermarked re-uploads and unedited imports — For You-ineligible verbatim

- - Extremely short clips / GIF-only videos / QR codes — named in the eligibility standards

- - Like-begging content — measured −60% interactions plus bait policy

- - 0%-volume trending-sound trick — audio analysis reads actual sound; the trick does nothing

- - Delete-and-repost accounts — duplicate detection flags reuploads; erratic view patterns on such accounts are artifacts

- - Low-effort AI slop — dies on quality/originality grounds; note AI content per se is labeled, not throttled, and can go enormous — judge the content, not the tool

+ Cluster by the shared creative device and viewer problem, not merely a repeated tag. Classify sounds, formats/memes, challenges, effects or topics as editorial categories; these are not a claim about the current UI tab names (directional).

- ### Saturation detection

+ | Dimension | Inspect | Don't infer |

+ |---|---|---|

+ | Momentum | Comparable chart snapshots and new relevant executions | A universal lifecycle from the age of a top post |

+ | Relative attention | Views/follower context and the account's normal range | Causal recommendation strength or genuine engagement |

+ | Discussion quality | Comments with use cases, objections or specific requests | Buyer intent from comment count alone |

+ | Relevance | Fit to the supplied audience and proof the author can show | Product demand from global popularity |

+ | Original contribution | A niche-specific demonstration, evidence or substantive response | Permission to copy an execution |

+ | Production readiness | Available footage, rights and usable sound | That a trend ranking grants a commercial license |

- A sound/format is saturated when: Creative Center rank has plateaued or fallen, 5+ near-identical executions cluster in the harvest, and the youngest breakout is older than ~5 days. The feed's same-sound diversification means late entrants compete for rationed slots. Recommend saturated trends only with a hard niche-twist, and label them as such in the brief.

+ Compute views/follower or comments/views only when visible counts and denominators are usable; disclose rounding and observation time. Compare similar account contexts and post ages. Never use length multiplied by views as a completion proxy; public views do not reveal completion.

+ Use ordinal decisions: recommend, watch, or reject. Recommend when relevant evidence and a feasible original execution support the brief; watch when momentum or rights remain uncertain; reject when fit or source quality fails. Avoid fabricated score thresholds (directional).

+ ### Assess freshness without a deadline myth

+ Compare observed chart direction with recent independent executions. Explain disagreements instead of forcing a verdict. A falling chart with fresh niche-specific examples may still support a useful post; a high rank without recent evidence may need another observation (directional).

+ Call a candidate crowded when the sampled executions repeat the same device without a useful new angle. Report that as a sample-bound judgment. Don't convert it into a platform saturation penalty, use a fixed age cutoff, or claim the feed reserves a fixed number of same-sound slots (directional).

+ ### Apply originality and commercial gates

+ Name the author's proposed original layer and any missing asset permission. Reused material without creative edits can be For You-ineligible (official, Integrity and Authenticity, 2025-08). Re-verify the current regional rule when a recommendation depends on it.

+ Capture disclosures and possible rights issues as observations, not enforcement verdicts. Distinguish public visibility from eligibility to republish, advertise or monetize. Don't fill the harvest with duplicated penalty lists or infer an AI-label reach effect from high or low views.

+ For campaign seeding, prioritize relevant existing creator contributions over a scripted wave of identical comments (directional). Fake engagement and repetitive irrelevant posting are covered by TikTok's authenticity rules (official, Integrity and Authenticity, 2025-08).

+ ### Return a usable brief

+ Include the audience/job, coverage and access gaps, candidate clusters, source records, observed momentum, proposed original angle, hook bank derived from visible structures, commercial constraints and the next recheck condition. Keep exact quotes separate from adapted hooks, and mark a proposed hook as unpublished craft (directional).

+ For launch research, identify real objections, demonstrable use cases, relevant creators, and which source asset could support each recommendation. Don't invent customer outcomes or convert campaign case studies into forecast returns. Give the writer the evidence and its limitations; don't draft an entire launch calendar during a harvest.

## Examples

- ### Example 1: Sound classification driving the recommendation

+ ### A rising sound with incomplete clearance

- ```

- Sound: "<track>" — Creative Center songs chart #14 and rising (was #31 two days ago per memory), country=US, industry=tech

- Top videos: 4 of 6 under 5 days old; two micro accounts at 40x and 85x views/follower

- Class: sound-locked (audio joke, no format skeleton)

- Verdict: BUY THIS WEEK — rising pre-plateau, micro-driven. Original layer required: niche-specific scenario over the audio, not a lip-sync.

- ```

+ Illustrative brief, not live observations (directional):

- ### Example 2: Rejecting a polluted stat

+ - Scope: [market], [buyer situation], observed [timestamp].

+ - Evidence: [sound URL]; chart moved from [earlier rank] to [current rank] using [matching filters]. Recent niche examples: [post URLs].

+ - Interpretation: sound-dependent candidate; recent executions address the supplied audience, but the commercial-use status isn't verified.

+ - Decision: watch pending clearance. Proposed original layer: demonstrate the author's actual packing problem in the sound's pause; use owned speech if music permission remains unresolved.

- A harvested "insight" claims duets get "+40% interactions per TikTok's early-2025 data." No such TikTok publication exists — the figure traces to aggregator farms. The brief describes the duet mechanic and its strategic logic (borrowing an anchor video's audience for commentary) with no percentage attached, and the source-hygiene note logs the rejected claim to memory.

+ This separates chart evidence from rights and leaves missing observations explicit. It does not turn rank movement into a deadline.

+ ### High views without retention evidence

+ Illustrative brief, not live observations (directional):

+ Source: [post URL], with [displayed views] at [timestamp]. Duration: [displayed duration]. Comparable posts: [URLs and visible counts]. Comments include [verbatim relevant excerpt].

+ Conclusion: relative attention is observable; completion and sales are unknown. Recommend inspecting the opening and the visible demonstration. Do not call the duration/view combination a retention win.

## Checklist

- Before returning the TikTok section of the brief:

- - [ ] Creative Center harvested with country + industry filters set first; charts trajectory (rising/plateau) recorded, not just rank

- - [ ] At least one keyword-insights pull and one tag/sound page per candidate trend

- - [ ] Every cited post has handle, account band, views, comments, age, and URL

- - [ ] Views-to-follower and comments-to-views computed; completion proxy noted for length outliers

- - [ ] Every trend classified sound-locked vs format-locked, with the matching window in the brief

- - [ ] Every recommended angle names its original layer (For You originality gate)

- - [ ] The 10-day lifespan applied — nothing older recommended as "live"

- - [ ] No figure in the brief without a named primary source or harvest evidence

- - [ ] Chart states + sound trajectories logged to memory for next-session velocity

- - [ ] All background harvest tabs closed before returning the brief

+ - [ ] Market, audience, buyer situation and commercial scope are explicit.

+ - [ ] Live entry points and resolved URLs are recorded; missing or renamed surfaces and access gaps are disclosed.

+ - [ ] Chart comparisons use matching filters and known observation times.

+ - [ ] Every quoted hook and number traces to a source record; absent data remains unknown.

+ - [ ] Organic posts, paid examples and generated suggestions are distinguished.

+ - [ ] Relative metrics have usable denominators; no completion proxy, fixed trend expiry or invented scoring threshold appears.

+ - [ ] Each recommendation names its original contribution, evidence, rights status and recheck condition.

+ - [ ] The brief contains relevant objections and usable hook structures without invented author experience.

+ - [ ] Commercial music and regional policy checks are identified where needed; no popularity-to-permission inference remains.

+ - [ ] Reusable observations are returned or stored only as authorized; research tabs are closed.

- ## Composition / References

+ ## References

- - Pairs with `social-tiktok` (content domain) for drafting once the brief is in hand — same mechanics, applied to creation.

- - How TikTok recommends videos (official): https://newsroom.tiktok.com/en-us/how-tiktok-recommends-videos-for-you

- - For You eligibility standards: https://www.tiktok.com/community-guidelines/en/fyf-standards

- - Creative Center: https://ads.tiktok.com/business/creativecenter — free, public; Symphony Assistant and product insights need login

- - Measured anchors: Metricool 2026 (n=2.31M posts — the 10-day lifespan, 72.7% FYP share), Socialinsider 2026, Buffer 2026. Lifecycle day-ranges are deliberately absent — use in-product chart trajectory; re-validate the eligibility standards and Creator Rewards terms periodically.

- - Use the agent's universal output schema; this skill only supplies the parameters that go into it.

+ - [TikTok One Creative Suite](https://ads.tiktok.com/creative/creativeCenter), undated, verified 2026-09.

+ - [SMB Playbook](https://ads.tiktok.com/business/library/Global_SMB_Creative_Playbook.pdf), undated, verified 2026-09.

+ - [Content Suite announcement](https://ads.tiktok.com/business/en/blog/content-suite-creator-ugc-library?redirected=1), 2026-06-22; [Product Preview](https://ads.tiktok.com/business/en/blog/tiktok-product-preview), 2026-07-28.

+ - [TikTok recommendation explainer](https://newsroom.tiktok.com/how-tiktok-recommends-videos-for-you?lang=en), 2020-06-18; [Integrity and Authenticity](https://www.tiktok.com/community-guidelines/en/integrity-authenticity), 2025-08-14.

+ Re-validate when: URLs redirect or navigation/filter labels change; login or regional access changes; music permissions change; recommendation or authenticity guidance changes.

+ Validated: 2026-09