Florence · {{version}}
Your AI Chief Data Analyst
for Amazon.
Florence runs your CRO loop end-to-end inside Claude Cowork. Mines reviews. Watches competitors. Generates image variants. Validates them with real shoppers. Logs every project in a tracker you own. About {{install-time}} to wire her up; runs forever.
Why Florence — AI slop vs. connected systems
Most CRO today is a stack of one-off prompts and disconnected tools. Midjourney for images. ChatGPT for copy. A spreadsheet for tracking. A Notion page for research. None of them know your brand. None of them remember last week. Every time you start, you start from zero.
Florence is the opposite. She's a connected system — your brand context, your products, your goals, your voice rules, all in one brain that travels with every task. She's the perfect example of scalable systems you own: build a workflow, integrate it with your stack, automate, train on your own brand. Then change it. It's yours.
Own your automations
Florence is yours. The skills, the brain, the artifacts. Fork her, modify her, take her with you. No subscription lock-in.
Own your data
Your brain.json sits on your laptop. The reference library is in your Project Files. Nothing leaves your Cowork instance unless you tell it to.
Scalable
Built using best-in-class CRO methodology + ProductPinion's test library + Higgsfield's image gen. Add your own MCPs. Connect your own tools. She grows with you.
How she works
Florence is built on the shoulders of:
- 100+ hours of CRO methodology — distilled from Keplo's client work + Matt Kostan's 2026 ProductPinion interview
- 17 ProductPinion test templates — calibration data for what actually lifts CTR / CVR
- Three creative-director references — main image (1:1), listing slots (1:1), A+ Content (16:9), each with a 13-step canonical flow
- 52-tactic main-image library — every variant Florence renders cites a specific tactic by number
- Claude Cowork as the runtime — plugs into Slack, Notion, Calendar, ProductPinion, Higgsfield, your own MCPs (custom or official)
How that translates: when you type optimize-listing B07X, Florence chains 5 SellerApp endpoints, triangulates objections from reviews + Rufus, scores the keyword pool against the metric-to-action quadrant matrix, cites the exact tactic from the library — and emits a portable research artifact you can hand to a designer. Your brain stays current; the next skill picks up where this one left off.
What Florence does — example skills
optimize-listing B07X
Fast CRO audit (~30s). 5 SellerApp endpoints chained, 3 ranked actions with 40/30/15/15 scores, citations to the reference library.
shopper-interrogator B07X
Full objection-mining loop. Reviews → Rufus → Pinion polls → image blueprint. The headline workshop demo.
render main image for B07X
5 image variants in ~90s. Pulls reference image, picks 5 enhancement techniques, generates at 1:1, scores each on 4 axes, iterates failures up to 3 attempts.
pinion launch image split
Validate with real shoppers. ProductPinion test panel, calibration-cited results, verbatim shopper quotes.
idea "test the no-grease angle"
Capture hypothesis. Surfaces on the cockpit Ideas tab. Promote to a project when you're ready to act.
today
Daily brief. Priority pick + signals + "what I'd do today" line. Schedule via Cowork Routine for 8am delivery.
Today's workshop · AI slop vs. connected systems
Florence is one of four sister tracks:
LoRAs primer — fine-tune image-gen models on your own products (Matt)
Ethical Prompt Injection skill — bias your shopper's own LLM toward your brand (Matt)
Systems Consultant Agent — interview agent that maps your custom system roadmap (Dorian)
Florence — the connected CRO system you can take home today (Matt + Dorian)
Type
ready when you are.
~{{install-time}} from here to a fully-wired Florence.
Brand → MCPs → Products → Voice → Goals → Done.