AI Workflows — Team Training

Three AI workflows Sean recorded to teach the team: ad-data analysis, business-strategy prompting, and AI UGC production. Each links its Loom — the videos show the actual screens and outputs.

1. Ad performance analysis (FB leads + appointments)

📹 Watch: https://www.loom.com/share/4f70a3a26e804bc6b78c720f07f66803

Turns the messy lead-tracking sheet (campaign/ad set/ad per lead) into pattern analysis and keep/kill recommendations.

1. Open the client's lead tracking sheet. Pick the start row for the analysis window (e.g., first lead of the month, or the row after your last campaign change).
2. File → Download → CSV.
3. Take the analysis prompt from the prompt pack → set "start analysis at row X" at the bottom.
4. Drop prompt + CSV into the AI and read the result: per-creative booking rates (e.g., "David vs Goliath: 24 leads, 11 booked, 45%"), ad-set-level patterns (same geo/age across sets → it's creative, not targeting), CBO vs ABO split, watch-thresholds.
5. Model choice: Claude. In Sean's three-way test, ChatGPT and Gemini both recommended cutting budget on a result Sean was happy with; Claude said "watch it" — the recommendation quality is why Claude is the default here (consistent with Media Buying Audit & Optimization Playbook).

2. Business strategy prompting (the Air Synergy example)

📹 Watch: https://www.loom.com/share/112cac366f9f4aba9cbc91f140850d34

How to get maximum strategic intelligence out of AI for a real business problem — taught via the March 2026 Air Synergy situation (client pushing for a pay-per-close offer we don't want to give).

1. Start from the business-strategy system prompt in the prompt pack (it forces source-citing, search-over-guessing, strategic framing).
2. Always pick the smartest model tier available (Pro/max-class), never the fast default.
3. Dump maximal context: the client notes, what the client said, what the CSM said, our criteria, quotes, PDFs, screenshots — everything.
4. Watch your framing. First attempt got refused because the prompt described the tactic in "bait and switch" terms; removing that terminology (and framing the actual, honest goal) produced an extremely useful answer. If the model balks, check whether your wording mischaracterizes the intent.
5. Output → hand to the owner of the problem. The Air Synergy answer became the playbook: make it a math conversation, be consultative not defensive — at $350/appt and a 40% close rate they acquire a customer for ~$875; the pay-per-close alternative is a $2K flat risk-premium fee, priced that way because we'd carry the close risk. Framed that way, the client talks themselves back into pay-per-appointment.

3. AI UGC production (Higgsfield)

📹 Watch: https://www.loom.com/share/5e6dce34ce384bd5b48890def6a02e3b

The 80/20 of Higgsfield: character image → talking clips → CapCut assembly. Prompt templates live in the shared doc linked in #media-buying-communication.

1. Character generation — do it in Google Flow, not Higgsfield (Flow is effectively free; save Higgsfield credits for video). Inputs: client logo, uniform details if known (e.g., black collared shirt). Output: 2K, 9:16. Batch until it looks right. - Selection criteria from the test project: in-vehicle shots read as authentic (steering wheel visible), logo accuracy matters, and older/bearded characters carry more authority than young-looking ones.
2. Video — Kling 3.0 with the character image as the start frame. Settings: multi-shot OFF, enhance OFF, audio ON, 1080p. (Motion-control variant = it mimics a video of you talking; a niche tool, not the default. Seed Dance lacks audio.)
3. Chunk the script into 7–8 second clips — longer single generations degrade toward the end. The chunking prompt splits the script for you.
4. Expect dead air at both ends of each clip (that's normal — it's how real people film).
5. CapCut: trim the dead air, hard-cut clip to clip. Jump cuts are native to UGC style — they add realism.
6. Export one file → feed the standard formats (split screen, TikTok explainer).
7. Recurring prompt ingredients (already in the template): handheld iPhone style, subtle wobble, no studio look, no plastic skin, natural lip movement / lip-sync emphasis. You only paste in the script.