Creator Sourcing

Creator Sourcing

Turns a niche into a ranked, contactable creator list plus a dashboard.

Requires the ScrapeCreators MCP connector. Confirm it with a keyword ToolSearch (e.g. instagram profile scrape) before promising anything; if no tools come back, say the connector isn't attached and stop.

Read reference/gotchas.md before writing any metric code. Every rule in it comes from a real failure, and several produce confident, plausible, wrong numbers when ignored.

Inputs

Collect these up front. Ask only for what's missing and genuinely changes the run; otherwise use the defaults.

Also worth settling early: mean or median engagement for the ranking (see gotchas §3), and whether the contact stage covers all results or just the top N.

Pipeline

Set CACHE to the session tool-results directory — large responses spill there and every script reads them from disk:

CACHE="$(dirname "$CLAUDE_SCRATCHPAD")/tool-results"   # or locate it from a tool result path

1 · Discover — three independent ways

Run all searches in one parallel batch. Three methods matter because each surfaces creators the others miss, and anyone found by 2+ is a stronger signal.

• profile/bio search × each keyword — v1_instagram_search_profiles
• hashtag search × each hashtag — v1_instagram_search_hashtag
• reels/video search × each phrase — v2_instagram_reels_search

Profile-search results usually return inline; write them to scratch/profile-search.json as {"<query>": [{username,url,follower_count,biography}]}. Hashtag and reels results spill to $CACHE on their own.

python3 scripts/collect_pool.py --cache-dir "$CACHE" \
  --inline scratch/profile-search.json \
  --reels-queries "phrase one,phrase two" \
  --out output/raw-pool.json

Check the reels label output it prints before continuing (gotchas §5).

2 · Enrich

One v1_instagram_profile call per handle with trim=true. It returns the last 12 posts inline, so this covers profile and posts in a single call. Batch ~16 at a time. Handles that fail with an SSE parse error need the fallback ladder in gotchas §4 — retry, then v1_instagram_basic_profile by userId, then record as unavailable.

python3 scripts/enrich.py --cache-dir "$CACHE" --pool output/raw-pool.json \
  --out output/enriched.json \
  --min-followers 5000 --max-followers 500000 \
  --max-age-days 60 --limit 40

Computes both mean and median engagement, drops creators silent past --max-age-days, applies the follower range, ranks by engagement rate, and writes output/run-meta.json with everything dropped and why. Read its stdout — it reports null-engagement creators and viral skew.

3 · Contact paths

Public sources only: the creator's own bio and their own link-in-bio page.

Bio emails are extracted automatically. Link-in-bio pages return inline, so resolve them yourself and hand over the contact fields:

• v1_linktree (returns an email_address field), v1_komi, v1_pillar, v1_linkbio, v1_linkme
• regex the whole payload for emails — one run found an address in Linktree's description rather than email_address (gotchas §10)

Write scratch/linkinbio.json keyed by handle (schema in scripts/contacts.py), then:

python3 scripts/contacts.py --enriched output/enriched.json \
  --linkinbio scratch/linkinbio.json --cache-dir "$CACHE" --top 40

--cache-dir also picks each creator's own external_url out of the cached profile, which is often the only contact path for accounts with no bio email.

Agency/management addresses get tagged automatically — flag them in your summary, since the outreach differs.

4 · Dashboard

python3 scripts/build_dashboard.py --enriched output/enriched.json \
  --meta output/run-meta.json --out output/creator-pool.html \
  --title "Creator Pool" --platform instagram --niche "skincare" \
  --captured "13 Aug 2026"

Self-contained HTML, no network calls. Verify before publishing: open it with the browser tools, exercise the filters against expected counts, check dark mode and mobile overflow. Then publish with the Artifact tool.

To update an existing dashboard, republish the same file path so it keeps its URL.

Reporting back

State plainly:

• how many were sourced, and how many survived each filter
• who was dropped and why — stale, out of range, or API-unavailable. A creator missing from a dashboard is indistinguishable from one never found.
• the mean/median split if any creator is viral-skewed, with the worst example
• which contacts are agency rather than personal
• credits spent

Do not present the pool as vetted for brand fit. Discovery matches captions as well as bios, so brands and off-topic accounts will be in there; say so.

Scope

This sources public professional data for partnership outreach. Keep to public bios and creators' own link-in-bio pages. Don't compile personal information beyond a business contact path, and don't send anything on the user's behalf without explicit approval.