Creator Discovery Router
Use this skill when the user wants to:
- find creators or influencers for outreach
- shortlist KOL / KOC candidates
- find mid-tier or micro creators in a niche
- find creators under follower constraints such as
- find creators by audience fit, not only by topic keywords
- compare discovery routes across TikTok, Instagram, and X
This skill is a routing layer. It should not be the primary collector.
It decides:
- what kind of discovery problem this is
- which platform skill should collect first
- whether to start from handles, content, or creator graph expansion
- when to hand off to
skills/50-publishing/creator-outreach
Follow shared release-shell and research rules in:
- release-shell rules
- research preferences
Execution Rules
Inside the product-shell runtime, keep this skill narrow and fast:
- this skill is route selection only, not a collection workflow
- do not enter plan mode
- do not spawn Agent/Task subagents
- do not use just to manage simple routing work
- do not use or any other discovery step to search for this skill's own reference files
- do not use Bash, , , , or similar shell exploration to inspect skill directories or references
- read only the directly relevant reference files listed in this skill with the tool
- treat the reference list in this file as explicit. If you need one reference, read that exact file path directly
- after choosing the route, either:
- explain the route briefly in user-facing language and hand off to the downstream platform skill
- or ask one short clarification question if the route is genuinely ambiguous
- do not end the turn immediately after only reading a reference file
- do not stop at route diagnosis; continue into the downstream skill in the same turn whenever the route is already clear
If the request already clearly maps to one route, do not stop for extra exploration.
Same-Turn Handoff Rule
For a clear creator-discovery request, this skill must complete three things in the same turn:
- normalize the request into a brief internally
- choose the discovery route
- hand off to the downstream platform skill
Do not stop after step 1 or step 2.
If the request already names a platform and the route is clear, do not return a route-only answer.
Default Handoff For TikTok Discovery
When the request is about finding TikTok creators/KOLs with topical fit plus a follower-band preference:
- default to
- explain the plan in plain business language
- hand off directly to
skills/20-research/tiktok-research
For requests like:
找 20 个 TikTok 护肤仪 KOL，优先微中腰部账号
do not pause after reading
. Route to
skills/20-research/tiktok-research
immediately unless a missing detail would genuinely change the route.
Do not treat a missing local reference file as a reason to stop or fail this handoff.
If the route is already clear from the user request, skip extra reference reads and continue into
skills/20-research/tiktok-research
.
User-Facing Explanation Rule
When explaining the plan to the user before collection, do not expose internal route labels such as:
Use business language that a marketer can understand.
Good examples:
- "我会先从最近真的在发这类内容的人里找，再补他们的主页和联系方式。"
- "我会先从竞品已经合作过的人和相关内容里拉一批候选，再筛掉不适合联系的账号。"
- "我会先用你现有的达人名单补资料，再按内容、互动和合作价值做一轮筛选。"
- "如果平台后台已经有达人池，我会先用那个做第一轮过滤，再补公开资料。"
Avoid explanations like:
- "我会先走 content-first"
- "这次更适合 graph-first"
- "我准备 mixed route"
Internal route labels are for system reasoning, not for user-facing communication.
Read these references before implementation:
- research preferences
skills/10-routing/creator-discovery-router/references/brief-schema.md
skills/10-routing/creator-discovery-router/references/routing-modes.md
skills/10-routing/creator-discovery-router/references/iteration-loop.md
skills/10-routing/creator-discovery-router/references/candidate-schema.md
skills/10-routing/creator-discovery-router/references/instagram-candidate-mapping.md
skills/10-routing/creator-discovery-router/references/x-candidate-mapping.md
Core Rule
Do not treat all "find creators" requests as keyword-based profile search.
First extract the real constraints:
- platform
- follower range
- recall range
- topic or niche
- audience
- geo or language
- creator type
- recency / activity
- contactability
Then choose the route.
Default assumption:
- do not use the user's target follower band as the first-pass recall band when the platform search is noisy
- use a wider recall band first, then tighten in shortlist scoring
Discovery Modes
Use when the user already has:
- seed creators
- competitor handles
- account usernames
- brand watchlists
Route:
- collect profiles from platform skills
- enrich recent content if needed
- rank and shortlist
- hand off to if partnership prep is needed
Use when the user wants:
- creators in a niche
- creators under a follower band
- real creators who are actively posting a topic
- partnership candidates based on what they post, not only what they claim in bio
Route:
- collect topic-relevant videos / posts / reels first
- extract authors from the content set
- enrich author profiles
- classify creator type
- filter by follower range and creator fit
- shortlist and hand off
This should be the default when the user asks for combinations like:
small study creators who post productivity workflows
mid-tier creators with overseas student audiences
Use when the best creators are unlikely to be found by simple search ranking.
Examples:
- micro creators
- local-language creators
- creators in a narrow subculture
- creators around a specific seed account or hashtag cluster
Route:
- start from one or more seed creators, hashtags, or posts
- expand through related accounts, tagged mentions, repeated collaborators, or creator clusters
- enrich candidate profiles
- filter and shortlist
Use when discovery needs multiple passes.
Example:
- content-first to get real active creators
- graph-first to expand around the best seeds
- handle-first to enrich the final shortlist
Routing Heuristics
If the user strongly cares about
:
- do not default to keyword-only profile search
- prefer or
- widen recall before tightening shortlist
If the user strongly cares about
:
- collect content before ranking creators
If the user strongly cares about
:
- do not trust bios alone
- use recent content patterns, language, and repeated framing as evidence
If the user strongly cares about
:
- classify candidates before final shortlist
- separate:
If the user already has seeds:
If the user wants contact-ready leads:
- collect first
- enrich second
- score third
- only then use
Platform Handoff
Use the narrowest useful platform skill:
- TikTok data ->
skills/20-research/tiktok-research
- Instagram accounts ->
skills/20-research/instagram-account-research
- X accounts ->
skills/20-research/x-research
- Outreach prep ->
skills/50-publishing/creator-outreach
Do not use public web search as the primary route for platform creator discovery unless platform collection is blocked.
Default Output
Return:
- chosen discovery mode
- why this route was chosen
- collection plan by platform
- filtering logic:
- recall range
- follower range
- relevance signals
- creator type rules
- exclusion rules
- whether the next skill should be collection, enrichment, or outreach
When returning creator candidates, normalize them into the shared candidate schema in
references/candidate-schema.md
.
Keep:
- a stable core schema across platforms
- optional fields for platform-specific or request-specific extensions
Do not let each platform return a completely different downstream shape if the outputs are meant to be merged.
Good Brief
Use the shared brief shape in
references/brief-schema.md
.
If the user gives a vague request, infer the smallest sufficient brief and proceed.
If the request is ambiguous in a way that changes the route, ask one short question.
Examples:
你是更希望我先从最近在发这类内容的人里找，还是先从你已有名单和竞品名单里找？
你这次更在意达人粉丝量级，还是更在意内容和目标受众的贴合度？
Failure Pattern To Avoid
Bad route:
- user asks for
- agent runs keyword-based user search
- agent filters follower counts afterward
- result quality is poor and sparse
Better route:
- recognize that follower band plus niche fit needs
- collect relevant content first
- extract and enrich authors
- classify creator type before final shortlist
- apply follower and fit filters after author expansion
When telling the user this plan, translate it into plain language:
- first find people who are already posting the right content
- then补主页和公开资料
- then按粉丝量、内容贴合度和账号类型筛 shortlist
Default Filtering Pattern
Unless the user explicitly requires strict first-pass filtering:
- use a wider recall band such as
- score relevance and creator type
- tighten to the target band such as
- return:
Do not collapse these into one list.
Iteration Rule
Do not assume one collection pass is enough.
Creator discovery is an iterative loop:
- collect a small but valid dataset
- evaluate the result quality
- diagnose where the failure or weakness is
- change one or two key variables
- run the next pass
Typical variables to change:
- route
- queries
- recall range
- creator type filters
- platform
- seed set
Do not change everything at once unless the current route is clearly invalid.
After each pass, decide whether the next step is:
continue on the same platform
change route on the same platform
Use
references/iteration-loop.md
for the evaluation and optimization checklist.