Talent Scout

Always-On Talent Scout

Turns one-shot discovery into a standing subscription habit. The user defines a scout profile once — one or more saved searches (brief-mode briefs, lookalike seeds, filter sets) — and this skill re-runs it on demand, diffs the results against the snapshot from the previous run (stored as a local state file in the workspace), and reports only what changed: new creators in the search space, creators who dropped out, and freshness movements. For agency casting leads running always-on programs and talent teams watching a niche.

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).

Honest scoping: there is no server-side saved-search or change-feed API. The scout is client-side — this skill stores result snapshots in the workspace and computes diffs locally. That also means the scout only "runs" when invoked. In harnesses with scheduled tasks (e.g. Cowork scheduled tasks), the user can schedule "run my talent scout" weekly and this skill pairs with that naturally — mention it when setting up a profile, but never depend on scheduling existing; on-demand runs are the baseline.

If the user has a one-off brief in hand, this is the wrong skill — use brief-to-shortlist. If they're watching competitor BRANDS rather than a creator search space, use competitor-watch.

Precision & filters for outreach (convention 14). When this list will feed an outreach step, default to precision over raw recall: search precision: "tight" with the hard-gate filters the brief supports (platform, niche, data_freshness_days, content_format, audience_country), and reserve broad plus heavy lookalike unioning for market sizing. Do not fan out dozens of lookalike calls to hit a volume target (each hop drifts from the seed), and trim the weak tail by each result row relative_fit (within-set fit, 1.0 = strongest) rather than padding to a round number. Canon: ${CLAUDE_PLUGIN_ROOT}/shared/conventions.md convention 14.

Inputs to collect

Setup (first run only):

• Scout profile name — e.g. "LATAM beauty mid-tiers". One workspace can hold several profiles; each gets its own state file.
• The saved searches (1–4 per profile) — each is either a brief-mode search (brief text + optional filters) or a lookalike search (seed creator or seed content). Capture them in exact search_creators argument form and show them back for confirmation; these arguments are frozen into the profile so every run is comparable.
• Result depth — limit per search, default 25 (thrifty: 15).
• Profile fan-out policy — default: profile only NEW creators each run (the diff is the product; profiling the whole result set every time wastes credits).

Re-run (every subsequent run):

• Which profile — only if more than one exists; never re-ask the saved criteria. The state file is the memory.
• Credit mode — default thorough; thrifty on the usual trigger phrases.

Flow

Step 0 — Load or create state. State lives at talent-scout/<profile-slug>.json in the workspace (create the directory if absent). Schema:

{
  "profile_name": "LATAM beauty mid-tiers",
  "created": "<ISO date>",
  "last_run": "<ISO date>",
  "searches": [ { "label": "...", "args": { /* frozen search_creators args */ } } ],
  "seen": { "<creator identifier>": { "first_seen": "<ISO date>", "last_seen": "<ISO date>", "label": "<which search>" } }
}

If no state file exists, this is a setup run: collect inputs, write the file, run the baseline (below), and tell the user the next run will produce diffs. On a baseline run the deliverable is the full result list labeled as baseline, not a diff — never fake a "new creators" framing on run one.

Step 1 — Re-run every saved search with its frozen arguments. One search_creators call per saved search, arguments verbatim from the state file, e.g.:

search_creators {
  "mode": "brief",
  "brief": "<frozen brief text>",
  "filters": { "country": "MX", "min_followers": 50000, "max_followers": 500000 },
  "limit": 25,
  "precision": "balanced"
}

A saved scout can freeze the GA hard-gated filters too (audience_country + min_audience_country_share for a standing audience-in-market watch, data_freshness_days, content_format) — capture them verbatim in the scout state like any other filter so every re-run applies the same gate.

Or a lookalike search:

search_creators {
  "mode": "lookalike",
  "seed_creator": { "type": "social_handle", "platform": "instagram", "handle": "<frozen seed>" },
  "limit": 25,
  "precision": "balanced"
}

Step 2 — Diff against seen. Classify each returned creator:

• NEW — not in seen → the headline of the report.
• Returning — in seen and returned again → update last_seen.
• Dropped out — in seen, was returned last run, absent this run → list briefly ("left the result set — ranking shift, not a verdict on the creator").

Step 3 — Profile fan-out on NEW creators only. get_creator_profile { "identifier": { "type": "<type from search result>", ... } } per new creator (exactly one identifier type per call). Thorough: all new creators. Thrifty: top 5 new per search by search-result signal; list the rest as "unprofiled — names only."

Credit estimate fires here if searches + expected new-creator profiles exceed ~30 credits (likely on baseline runs with 3–4 searches at limit 25 — state the estimate, offer thrifty, proceed).

Step 4 — Freshness Gate on every profiled creator per the shared module. Stale new arrivals go to "re-verify before pitch", never the headline list.

Step 5 — Update state and write the deliverable. Merge this run's results into seen, stamp last_run, save the file, then write the diff report.

Deliverable

# Talent Scout — <profile name>
_Run <date> · previous run <date or "baseline run — no diff yet"> · Creatorland Data_

## New in your search space since last run (<N>)

### <Name> — @<handle> (<platform>, <follower count>)
- **Entered via:** <which saved search matched>
- **Why they fit:** <1–2 sentences tied to the saved brief's language / seed similarity>
- **Audience-geo:** <top geos from profile>
- **Freshness:** fresh | aging (note)
<repeat per new creator; if zero: "No new creators this run — your saved space is stable. (That's signal too.)">

## Dropped out of the result set (<K>)
<names + which search; "ranking shift, not a verdict">

## Returning (stable presence)
<count only, e.g. "19 creators returned again across your 3 searches" — names on request>

## Re-verify before pitch
<stale new arrivals; or "None — all new arrivals cleared the freshness gate.">

## Scout profile on file
<the saved searches, labels + one-line summaries, so the user can ask to edit them>
> Tip: in a harness with scheduled tasks (e.g. Cowork), schedule "run my
> talent scout '<profile name>'" weekly to make this a true standing watch.

---
Data freshness: <N>/<M> profiled creators synced within the last sync window; <K> flagged for re-verification.
Provenance: Creatorland Data MCP · <S> saved searches re-run + <P> profiles · diff computed locally against <prior run date> snapshot · <date>.
Credits used this run: ~<N> (<S> searches ×2 + <P> profiles ×1).

Honesty rules

• The diff is client-side: "new" means new to the locally stored result snapshot, not "new to the platform" or "newly active." Say so in the report the first few runs.
• A baseline run is labeled baseline — never dress a first run up as a diff.
• "Dropped out" means left the ranked result set, nothing more — no churn or decline narrative.
• Thin-niche honesty (convention 8): if a saved search returns very few results, disclose and offer ONE labeled broadening step for the profile — never silently loosen frozen search arguments.
• Stale new arrivals never headline (Freshness Gate).
• No contact info anywhere in the report (convention 7); the affordance is the Creatorland Member flag / public profiles.

Credit footprint

thorough re-run: ~10–22 credits (2–4 searches ×2 + ~5–15 new-creator profiles ×1; quiet weeks cost as little as 4–8) · thrifty re-run: ~7–14 (same searches, ≤5 profiles per search) · baseline/setup run: ~25–55 if the full result set is profiled — the estimate fires and thrifty (names-only baseline, profile later) is offered. No market-intel calls in this skill.