Longlist Machine

Longlist Machine

Volume casting for always-on programs. Instead of one tight search, it runs the same brief through several phrasings, pools the results, dedupes, and tier-tags a 50–150-creator longlist the casting team works down themselves. Breadth and recall are the product here — curation happens downstream.

Read first: ${CLAUDE_PLUGIN_ROOT}/shared/conventions.md (tool schemas, credit prices, the nine conventions). This skill honors thrifty/thorough credit modes (${CLAUDE_PLUGIN_ROOT}/shared/credit-modes.md), Refusal Recovery (${CLAUDE_PLUGIN_ROOT}/shared/refusal-recovery.md), and the Freshness Gate (${CLAUDE_PLUGIN_ROOT}/shared/freshness-gate.md).

Want a curated, client-ready cut of 8–12 with full per-creator rationale? That's brief-to-shortlist. This skill optimizes for recall and volume — a working pool, not a finished deck.

Precision & filters for outreach (convention 14). This skill optimizes recall on purpose — right for a working pool, wrong for a list you will contact. The moment a cut of this longlist feeds an outreach step, switch to precision: re-cast that cut at precision: "tight" with the hard-gate filters the program supports (platform, niche, data_freshness_days, content_format, audience_country), and trim the low-relative_fit tail (within-set fit, 1.0 = strongest) rather than shipping the raw recall pool to contacts. Keep broad for the sizing pool, not the outreach cut. Canon: ${CLAUDE_PLUGIN_ROOT}/shared/conventions.md convention 14.

A sizing question is not a longlist run. If the user is really asking how many creators exist for a brief ("how big is the pool", "is there enough supply"), answer with count_creators first — a free, rounded headline count — instead of paying to build a 50–150-row longlist just to gauge size. search_creators (this skill's engine) returns a capped page, never a census. Build the longlist once the user actually wants the pool.

Inputs to collect

• The brief (required) — any written format; normalization is your job.
• Target longlist size — default 100; cap 150 (the search limit ceiling per call). Ask only if unstated and the word "longlist" alone doesn't imply a number.
• Phrasing axes (optional) — if the user knows the program has variants (e.g., "skincare vs makeup", "GRWM vs tutorial"), capture them; otherwise you generate the phrasing variants yourself in step 0.
• Tier tags wanted — default macro/mid/micro by follower band.
• Credit mode — default thorough; thrifty on trigger phrases. Longlist runs are inherently credit-heavy — flag this up front.

Flow

Step 0 — Generate phrasing variants (no tool call). From the brief, write 3–5 distinct brief phrasings that approach the same program from different angles (audience framing, content-style framing, niche-adjacent framing). The point is recall: different phrasings surface different creators. Show the variants so the user can prune/add.

Step 1 — Multi-phrasing search. One search_creators per phrasing, each at high count:

search_creators {
  "mode": "brief",
  "brief": "<phrasing variant N, ≤2000 chars>",
  "filters": {
    "platform": "<if the program names one>",
    "country": "<if scoped>",
    "niche": "<if clean>",
    "min_followers": <if a tier floor applies>,
    "max_followers": <if a tier ceiling applies>,
    "audience_country": "<only when the program targets an audience-in-market>",
    "data_freshness_days": <only when the program wants recently-active creators>,
    "content_format": "<personality_led | faceless_clip — only when the program implies a format>"
  },
  "limit": 150,
  "precision": "broad"
}

Use precision: "broad" for maximum recall (this is a longlist, not a precision cut). Thorough: 4–5 phrasings at limit 150. Thrifty: 2–3 phrasings, limit ≈ target size, precision "balanced".

Hold the GA hard-gated filters (audience_country + min_audience_country_share, data_freshness_days, content_format) constant across every phrasing when the program requires them — they filter the pool and add no rank, so applying them unevenly would skew the merged longlist. Advisory: add them only when the program truly calls for audience-in-market, recency, or a specific content format.

Step 2 — Merge + dedupe (no tool call). Pool all results; dedupe on creator identity (same creator from two phrasings = one row). Record how many phrasings each creator hit — a creator surfacing across multiple phrasings is a stronger, more robust match; that's a free ranking signal at zero extra cost.

Step 3 — Tier-tag (no tool call). Tag each creator macro/mid/micro by the follower band returned in the search result. No profile call needed for the tag — search results carry follower count.

Step 4 — Light enrichment (profiles). A longlist does NOT profile every row — that would be 100+ credits. Profile only:

• thorough: the top slice (e.g., top 20–30 by multi-phrasing hits) for the "verified head" of the longlist, plus enough to run the Freshness Gate on the head.
• thrifty: top 10 only, or skip enrichment entirely if the user just wants the raw pool — say which.

get_creator_profile { "identifier": { "type": "creatorland_user_id", "creatorland_user_id": "<id from search result>" } }

Credit estimate fires here — mandatory. A 5-phrasing thorough run with a 30-profile head is ~5×2 + 30 = ~40 credits; state it and offer thrifty before the run. Be explicit that the un-enriched tail of the longlist carries search-result data only (tier, follower count), not full profiles.

Step 5 — Freshness Gate on the enriched head only. Classify the profiled head fresh / aging / stale. The un-profiled tail is labeled "unverified — search-result data only" as a column, never presented as freshness-checked.

Step 6 — Thin-program honesty (convention 8). If even multi-phrasing recall stays thin for the niche, disclose and broaden one labeled step; mark the broadened additions as adjacent. Don't pad a longlist to hit a round number silently.

Deliverable

# Longlist — <program name>
_<L> creators, deduplicated across <P> brief phrasings · <date> · Creatorland Data_

## Phrasings used
1. <variant 1> · 2. <variant 2> · … <so the team can see the recall net cast>

## The longlist
| # | Creator | Tier | Followers | Phrasing hits | Status | Fit |
|---|---|---|---|---|---|---|
| 1 | @<handle> (<platform>) | mid | <count> | 4/5 | verified · fresh | exact |
| 2 | @<handle> (<platform>) | micro | <count> | 2/5 | unverified — search data only | exact |
| … | | | | | | adjacent (<broaden step>) |
<sorted by phrasing-hit count desc, then tier; "verified" rows are the
enriched head, "unverified" rows carry search-result data only>

## Verified head (enriched)
<for the profiled top slice: a compact note per creator — audience geo,
freshness, and a paid-plan avatar (`avatar { url, source }`, `null` on
free/demo) so the team can eyeball the most-vetted rows; initials when null>

## Re-verify before pitch
<stale creators among the enriched head; or "None among the verified head.">

## Caveats
- This is a working pool optimized for recall, not a curated cut — expect to
  work it down. The tail is search-result data only (tier + follower count),
  not freshness-checked.
- "Phrasing hits" is a robustness signal (more = surfaced across more angles),
  not a quality guarantee.
- <if broadened: which adjacent niches and why exact recall was thin>
- Rate context omitted by default for a longlist — run `brief-to-shortlist` or
  `one-number-rate` on the cut you choose.
- No creator contact info is included or available via this tool; route
  outreach through Creatorland connections or the creators' public profiles.

---
Data freshness: <N>/<H> enriched-head creators synced within the last sync window; <K> flagged for re-verification. Tail (<L−H>) is search-result data only.
Provenance: Creatorland Data MCP · <P> brief searches + <H> head profiles · <date>.
Credits used this run: ~<N> (<P> searches ×2 + <H> profiles ×1).

Honesty rules

• The un-enriched tail is always labeled "search-result data only" — never presented as profile-verified or freshness-checked.
• "Phrasing hits" is a recall-robustness signal, not a quality score.
• Thin niches broaden one disclosed step; broadened rows labeled adjacent (convention 8) — no padding to a round number.
• Freshness Gate applies to the enriched head; the tail's freshness is explicitly unknown.
• Never imply access to creator contact info (convention 7).

Credit footprint

thorough: ~40 credits (5 searches ×2 + ~30 head profiles ×1; +5 per refusal rung if a rate call is added on request) · thrifty: ~16 credits (3 searches ×2

• ~10 head profiles ×1), or ~6 credits raw-pool-only (3 searches ×2, no enrichment). Profiling the FULL longlist is deliberately avoided — it would be 100+ credits; the head-enrichment pattern keeps cost bounded.