Case study: 50 US baby-stroller KOLs in one Friday afternoon

The setup

The brand had a workspace already, with brand profile filled in:

• product_category: "Premium baby gear (DTC strollers)"
• target_audience_text: "US Gen-Z + millennial new parents, household income $80k+, urban or suburban-coastal"
• target_geographies: ["US"]
• exclusions_text: "no political content, no influencer-marketing-tool reviews, no gambling adjacency"
• Sender identity: name + title + email signature

Three reusable inputs, set once, that change how every downstream LLM call scores fit and signs emails.

The walkthrough — 65 minutes

2:00 — Wide-net discovery (5 min)

From Claude Desktop:

"Run a TikTok keyword scrape for 'stroller review', max_videos 200, and another for 'newmom musthaves', max_videos 100. Show me the job_ids."

Two scrape_tiktok_by_keyword calls, two job_ids returned. Operator goes to make a coffee.

2:15 — First triage (10 min)

Scrapes done. ~180 authors scraped, mixed quality.

"List the KOLs from those two jobs that are in the US and aren't brand accounts. Sort by follower count descending. Show me 30."

Claude calls list_kols with country=US + hide_sellers=true. Returns 27 actual creators (out of the 180 starting set — most of the noise was brand/store accounts the classifier flagged).

2:25 — Pick one good seed, expand (10 min)

Operator scans the 27, picks the first creator that looks like a strong fit — call her @sarah.stroller.

"Find 20 creators similar to @sarah.stroller. Show me the ones with co-mention score ≥ 10."

find_similar_creators returns 23. Operator filters for co-mention ≥ 10 → 14 high-confidence candidates (collab partners + cross-promoted creators in Sarah's actual orbit).

2:35 — Bulk enrich the union (5 min to queue, 15 min to finish)

Combine the 27 from the filtered list + 14 from the similar expansion → 41 unique handles. Add ~10 more from a spreadsheet the brand's email team had been collecting.

Paste into /kols/import, max_videos = 30, target list = "Fall stroller launch seeds". Submit. 51 jobs queue.

Operator takes another break. Apify chews through 51 lookups in ~15 minutes (parallel, 3-4 concurrent).

2:55 — Scoring + cold-outreach drafts (15 min)

All 51 enriched. Operator skims the list page: the brand profile is set, so every row shows a star rating in the compact card. Sorts by brand-fit ≥ 4. Down to 38 rows.

For each of those 38: click the dossier row → the Creator Overview is already cached (first dossier open after enrichment triggers it free), click "Cold outreach templates" in the header → drafts generated using workspace sender identity, cached on first generate.

Operator quick-edits the 5 highest-fit drafts for tone, leaves the rest as-is. (The first-pass drafts are usually fine; the polish budget is reserved for the top of the list.)

3:10 — Export to Notion (5 min)

From Claude:

"Create a Notion table with the 38 KOLs in 'Fall stroller launch seeds', columns: handle, followers, ER, brand-fit score, has_email, cold-email subject + body. One row per KOL."

Claude pulls from list_kols + the cached cold- outreach rows (no LLM call — they were cached during the 2:55 review), creates a database in Notion via the Notion MCP server. Done.

3:15 — Review + Slack handoff (5 min)

Operator posts the Notion link in the brand's marketing Slack with: "Reviewing Monday 10am. Feel free to comment in the row notes before then." Done.

Where the credits went

• 2 keyword scrapes × 1 credit each = 2 credits
• 51-handle bulk enrich × ~1 credit each (30 videos) = 51 credits
• 38 cold-outreach drafts × 1 credit = 38 credits
• ~5 Creator Overviews refresh-generated for the closest reads (the rest were on-first-paint, free for now) = 5 credits
• ~3 Deep Analyses for the top candidates × 5 credits = 15 credits

Total: 111 credits, plus the rough ~300 implicit credits the brand had already spent maintaining their workspace credit pool. Roughly $3.20 in actual underlying spend at our cost basis; ~$8 at retail.

What this used to look like

Before this release wave, the same outcome required:

• A VA spending 4-6 hours scrolling TikTok manually, screenshotting profiles into a Google Sheet
• A second VA / contractor manually clicking through bios for email addresses
• The marketer hand-writing 30+ cold emails because nobody shipped reliable AI drafts that knew the brand's voice
• Sunday-evening cleanup before the Monday review

Quote from one of the three operators we composited this from: "The bit that finally clicked was when I realised the brand profile and the sender identity meant I never had to babysit the drafts again."

What's not magic

Three honest limitations:

• Email reply rate is still your job. KOLens drafts personalise the hook; getting opened still depends on subject-line discipline and which inbox you send from.
• Brand-fit scoring is a recommendation, not a contract. The LLM is grounded but it's still a 1-5 estimate. Review the ones it flags 5-star.
• The classifier is heuristic, not perfect. A real creator with "shop my looks" in bio gets flagged seller-like. Audit the toggle on edge cases.