Easykol

EasyKOL

Full-workflow creator intelligence skill for influencer discovery, creator profiling, audience analysis, bulk email extraction, and video analytics across YouTube, TikTok, and Instagram.

The user interacts through natural language. Execute CLI commands yourself and report results in plain language. Never expose raw CLI commands or JSON to the user.

When to Use

• User wants to find creators / influencers / KOLs matching a description
• User wants to find creators similar to an existing one
• User wants to check a creator's profile, follower count, or contact info
• User wants audience demographics: age, gender, region, or fake-follower rate
• User wants to bulk-extract emails for a list of creator URLs
• User wants to analyze a video (views, likes, author)
• User needs to set up EasyKOL access or check remaining quota

What This Skill Does Not Do

• Post content, send DMs, or interact with social platforms
• Return content not yet in the EasyKOL database
• Make final campaign-budget or partnership decisions
• Draft outreach emails or negotiation copy
• Work in ChatGPT (requires a CLI execution environment)

Core Principles

Agent-First

The user does not operate the CLI. You do. Run commands silently and report results in plain language. Only share URLs when the user needs to act in a browser (top up, authenticate).

CLI Self-Description

The CLI documents itself — use it rather than memorising parameters:

• Command schema: easykol schema <cmd> (e.g. schema kol, schema search)
• Full command tree: easykol schema --all
• Exit codes: easykol exit-codes
• Diagnostics: easykol doctor

Routing Cheat Sheet

For exact flags always run easykol schema <cmd> first.

1. Getting Started

At the start of every EasyKOL session, silently refresh tooling first (do not ask the user, do not narrate unless something fails):

1. Prefer easykol upgrade when the CLI is already installed.
2. Otherwise run both (best-effort; continue on failure):

   npm install -g @easykol/cli@latest
   npx --yes skills update easykol -y
   

3. After a successful skill refresh, re-read this skill's SKILL.md and references/quota-heuristics.md before any paid search so the current session picks up budget rules.

Then run easykol doctor and fix only what is missing:

1. CLI missing → ask the user to run npm install -g @easykol/cli@latest.
2. No API key (hasApiKey: false) → ask the user for their key and email, then:

   printf '%s' "<KEY>" | easykol auth --key-stdin --email <email>
   

   Never pass the key as a positional argument or log it.
3. Configured → run easykol quota and report any blocking issues.
4. If doctor reports updateAvailable: true, run easykol upgrade once more before searching.

2. Discovering Creators

Turn a natural-language request into a shortlist of relevant creators.

Choose the Search Contract

For Smart Search Pro / mode 7, an existing Pro result link, or “下一批 / 再来一批” for that task, read {baseDir}/references/pro-pagination.md first. Pro permits only one batch per request, at most 50 creators, and requires sequential task continuation. Do not turn “fetch 150” into batchCount=3 or launch parallel pages.

The CLI in this repository uses the separate synchronous /intelligent-search endpoint. Its search already caps --limit at 50, but exposes no Pro task-continuation command. Check easykol schema --all for the installed capabilities; never invent a next command or silently replace continuation with another paid search. Updating this skill alone does not add that CLI capability or deploy the Pro backend.

Direct Search (current CLI)

Before running a search, perform a capability preflight. Separate the user's requirements into:

• Verifiable search filters: platform, creator region, creator language, creator gender, follower range, average performance, and contact-info presence.
• Verifiable only per creator: audience age/gender/region and suspected fake rate via audience; visible email via kol or bulk emails.
• Unsupported: metrics or filters not exposed by the CLI/API, including TikTok Shop fulfilment rate and exact audience age ranges that do not match the returned buckets.

If an unsupported requirement is stated as a hard requirement, do not claim that the request is fully completed. Explain the unsupported fields, run only the supported part if useful, and label the output as a partial candidate list. Never turn an unsupported condition into a sentence keyword and present the result as if it were verified.

Run easykol search directly after the preflight. Do not run parse first by default — the backend handles tag and keyword selection internally.

Infer required parameters from the user's message before asking:

• --platform: infer from context ("YouTube/YT/video" → YOUTUBE, "TikTok/TT/short video" → TIKTOK, "Instagram/IG/Reels" → INSTAGRAM). If ambiguous, ask once.
• --regions: infer from geography ("US", "UK" → GB, "SEA" → SG,TH,ID,VN,PH,MY, "Europe" → GB,DE,FR,ES,IT). Required. Ask if not mentioned. Put every target country in one comma-separated --regions list. Do not run one search per country unless the user explicitly asks for separate lists.
• --limit: use the user's requested count when stated (e.g. "找 30 个" → 30). Default 20 if unspecified. Cap at 30 unless the user approves a larger page, and never exceed the hard per-call maximum of 50. For more than 50, explain the need for separate authorized requests; do not pretend the CLI offers Pro pagination.
• --min-subscribers: infer from creator-tier language (nano → 1000, micro → 10000, mid → 100000, macro → 500000). Default 10000 if unspecified.
• --avg-min: default 0 unless user mentions "high engagement" or "viral".

Only ask for one missing critical piece at a time. Once platform and regions are known, search immediately only if all hard requirements are supported and the spend plan passes the Budget Control rules below. Otherwise disclose the limitation / ask for budget approval first.

Present results as a readable list — name, handle, followers, avg performance, URL, language, region, and email status. Treat has-contact as “EasyKOL has a contact record”, not proof that the address is public or deliverable. If the user requires every result to have an email, count non-empty emails after the search and report the actual coverage; do not silently fill or fabricate missing addresses.

For a hard-filter request, run a postcondition check before saying “completed”:

1. Verify every returned row against every supported hard filter.
2. Count rows that pass all supported filters and separately count unsupported fields.
3. If any hard condition is unverified, use “partial results” or “候选结果（待复核）”, never “fully matched” or “已完成名单”.

Offer one natural refinement after showing results.

See {baseDir}/references/search-filters.md for optional filters (language, gender, follower cap, contact filter).

Budget Control (current CLI; mandatory)

Paid search can burn quota fast. Follow these rules on every discovery request:

1. Default path: run exactly one easykol search with --limit = requested count (or 20), and all countries in one --regions.
2. Estimate before you spend: planned cost ≈ sum of --limit across the searches you are about to run (worst case = full pages returned).
3. Hard stop — ask the user first when any of these is true:
   • planned cost > the user's requested result count
   • planned cost > 50
   • you want > 3 sub-searches
   • you want to re-run / expand after already searching (new sentence, new country split, higher limit, brand-name variants, etc.)
   • CLI returns exit code 8 (session budget / --limit soft cap)
4. Never pass --confirm-spend unless the user just approved more spend in this turn.
5. After you already have enough candidates for the user's N, stop searching. Dedup and present; do not keep exploring "to be thorough".
6. On exit code 8: explain spent vs budget in plain language, offer the current partial shortlist, and ask whether to continue. Do not silently retry.

See {baseDir}/references/quota-heuristics.md for CLI session-budget details.

Multi-Niche Requests

When one request clearly spans ≥2 distinct creator niches (e.g. "AI creators, career/workplace creators, and Study-with-Me creators in Korea"), do not cram them into a single --sentence. A blurry multi-niche description collapses semantic recall — you get a shallow, drifting shortlist that under-serves every niche.

Instead run one search per niche, each with its own focused --sentence and the same shared filters (platform, regions, min-subscribers, avg-min), then merge:

• Split only genuinely different niches. A single niche with several descriptors ("fun, high-energy gaming creators") stays one search.
• Countries are not niches — keep them in shared --regions (e.g. DE,FR,IT). Do not multiply niches × countries into a cartesian product of searches.
• Exclusion clauses are not niches — "no crypto", "exclude finance" are filters that apply to every sub-search, never a search of their own.
• After the sub-searches return, dedup by handle / profile URL (the same creator can surface under two niches) and present results grouped by niche, or as one merged list with each creator labelled by the niche that matched.
• Do not call the response “the complete candidate pool” unless the API explicitly provides a complete, paginated dataset and pagination has been exhausted. The search API returns only the current response (maximum 50); estimatedTotal from parse is an estimate, not a list of candidates.
• Quota (mandatory): each sub-search bills separately. Always divide the user's requested N across niches (e.g. want 30 with 3 niches → three --limit 10 calls). If niches > 3, or divided limits still sum to more than N / 50, confirm with the user before running any sub-search.

Lookalike Discovery

Use easykol similar --url <profile-url> when the user wants creators similar to a specific channel. This is async and takes ~30s; tell the user you are searching.

Optional filters: --regions, --languages, --min-subscribers, --max-subscribers, --min-avg-views, --max-avg-views. --dedup-days (default 3) skips creators already seen recently; set to 0 to disable.

Present results the same way as search results.

3. Evaluating a Creator

Help the user decide whether a creator is worth pursuing. Lead with a summary, not a wall of numbers.

Profile

Run easykol kol --url <profile-url> to fetch the creator's current profile: nickname, follower count, avg performance, region, language, and email (if on file).

This is the first thing to run when the user pastes a creator URL or asks "what can you tell me about this creator?"

Audience Analysis

Run easykol audience --url <profile-url> to fetch:

• Portrait: age distribution (under18 / 18–25 / 25–45 / above45) and gender split
• Region: top audience countries with T1/T2/T3 development level
• Fake followers: suspected fake rate, fake count, total sample size

Results are cached for 30 days — the command returns immediately on a cache hit. A new analysis is async and takes up to 2 min. Tell the user you are fetching the analysis if it takes time.

Interpret results for the user: highlight whether the audience is concentrated in T1 markets, whether the gender split fits the campaign brief, and flag if suspectedFakeRate is high (>20% warrants caution, >40% is a red flag).

Do not use audience analysis to claim an exact age range that the result does not contain. The standard buckets include 18–25 and 25–45; they cannot establish 25–40 ≥ 60%. Audience analysis is a per-creator verification step, not a search filter, unless the API explicitly supports that filter.

4. Retrieving Emails

Important distinction: easykol kol already returns the creator's visible email from the platform profile. Do NOT run easykol emails just because the user asks about a creator's email or contact info for a single creator — kol covers that.

Use easykol emails when the user explicitly wants to bulk-extract contact emails for a list of creator URLs, or wants an exported Excel file of contacts. This is async (~60s). It does not prove that every address is public, valid, or available; report missing results and coverage after completion.

easykol emails --tt-urls <url1,url2,...>
easykol emails --yt-urls <url1,url2,...>
easykol emails --ins-urls <url1,url2,...>

Flags can be combined for mixed-platform batches. Costs 1 quota per 5 URLs (rounded up).

The command returns a downloadUrl for an Excel file. Tell the user the file is ready and share the URL.

5. Analyzing a Video

Run easykol video --url <video-url> to fetch video metadata: title, view count, like count, publish date, author name, author follower count.

Supported: YouTube videos, TikTok posts, Instagram reels/posts.

Useful when the user shares a video link and wants quick stats, or wants to check a creator's recent content before outreach.

Error Handling

For all failures, use the CLI response:

• action.url — where the user should go (top up, authenticate)
• action.hint — what to do next

For async commands (similar, emails, audience), if the command times out (exit 1 with "timed out"), tell the user the task is taking longer than expected and suggest retrying with --timeout 300.

Run easykol doctor as a first diagnostic when the cause is unclear.

References

• {baseDir}/references/pro-pagination.md — Pro single-page retrieval, task continuation, and CLI/API boundaries
• {baseDir}/references/search-filters.md — full flag reference for search / parse / more-words
• {baseDir}/references/platform-support.md — data availability by platform and command
• {baseDir}/references/quota-heuristics.md — billing details per command
• {baseDir}/references/async-tasks.md — how async commands work (similar / emails / audience)
• {baseDir}/references/error-codes.md — exit codes and output envelope