01 · The idea
What this is
Luca's ask: "Custom videos for people we are trying to prospect to. It needs to feel like a movie trailer or TV ad specifically about our value proposition and the value to the person we are reaching out to as well as their company."
We give the system a LinkedIn URL. It researches the person, their company, and their tech; writes a grounded 45-60 second trailer script; generates voiceover and one or two cinematic AI shots; assembles a polished branded video with their logo, live website, and real facts. Every video passes human approval before generation money is spent, and again before anything reaches a prospect. Delivery is a personalized landing page linked from the outreach email.
Fully automated per video, with no manual video building. The one-time human craft is designing the master trailer template; after that, every video is machine-assembled fills. We already run image-generation automations without human intervention on this same stack, so this is assembly of proven parts, not invention.
First target: Paul Mather, Managing Director, Marlowe Fire & Security. Our research already caught something important about his company (§6).
02 · The ask
What we need from you before video #1
What is our value proposition for this campaign? One clear promise, one sentence, aimed at an operations / BD director.
Every script currently holds a {OUR_VALUE_PROP} placeholder. The video's 0:15-0:30 beat is this sentence. Nobody can write it but you.
What's the CTA? 15-minute call? Calendar link? Reply to the email?
It's the last ten seconds of every video and the button on every landing page.
Volume expectation for months 1-2? We've planned for 10-20 per month.
Determines how much automation hardening to do now versus later.
Who signs off the creative bar? Proposed: Kyle approves the master template once; Josh approves each video.
The #1 kill-risk is a video that looks like generic motion graphics instead of a trailer. Someone must own "this looks cinematic."
Delivery channel? Email with video thumbnail + landing-page link (recommended), or LinkedIn DM with native video.
Changes the delivery integration we wire up.
Everything below is our recommendation. Tell us where you disagree.
03 · The product
The video itself
Fixed 45-60s structure; only the personalized content changes per prospect. This is what keeps every video polished and repeatable:
| Time | Beat | Example: Paul Mather |
|---|---|---|
| 0:00-0:05 | Hook. Cinematic shot + punchy line naming their world; their logo resolves in | "In UK fire & security, compliance never sleeps, and neither do the risks." |
| 0:05-0:15 | Problem. The specific tension for their role, grounded in real research | Engineer teams, 24/7 monitoring, data in a dozen systems |
| 0:15-0:30 | Value prop. One clear promise, shown not told | {OUR_VALUE_PROP} ← you supply this |
| 0:30-0:45 | Their company, specifically. Name them, show their live site, cite a real fact | The Marlowe → Mitie Fire & Security transition itself |
| 0:45-0:55 | Soft CTA. One ask | "Worth a 15-minute look?" + link |
Production values: cinematic music bed with a hit on the CTA, kinetic typography (max six words on screen), their brand color pulled from their site, one consistent AI narrator voice as our signature, cuts on the beat. No AI human faces or avatars. Research consistently shows they damage trust in cold outreach. The "wow" comes from one or two short atmospheric AI shots plus real brand assets.
04 · The system
How it works
- Intake: LinkedIn URL + prospect name (required: it's the wrong-person tripwire) + optional angle and campaign tag.
- Research & prep (auto, minutes, pennies): scrape profile + company + website (no LinkedIn login, zero account risk) → research brief with every claim source-tagged; unknowns stay blank, never guessed → scene-by-scene plan → automated fact-checker → asset auto-acquisition: logo, website screenshot (cookie banners stripped, redirects followed), brand color. Output: one review card. The info, the plan, and an asset checklist where every item is ready low-confidence or missing, with a manual-upload slot as last resort.
- Gate 1, plan approval: we fix flags, edit lines, press Generate Video. Nothing expensive has been spent yet; iterating here is free.
- Generate & render (auto, 5-15 min): per-scene voiceover → 1-2 AI hero shots (auto-archived to our storage before provider links expire) → template render → notification.
- Gate 2, video approval: Josh reviews and tweaks. Text edits re-render free from cached assets; one voiceover line costs cents; a hero re-roll is an explicit, clearly-priced button; every take is kept in a gallery.
- Deliver (auto): personalized landing page + link for the outreach email; per-video cost logged.
Anti-hallucination is architecture, not a prompt wish. The AI can only use facts we scraped; every claim carries its source; the verifier and Gate 1 sit before any generation money is spent; nothing reaches a prospect without approval at Gate 2.
Phase 2, full auto (only after Phase 1 is proven): Gate 1 auto-passes videos with zero flags and zero fact-check violations; anything else still stops for a human. Graduation is data-driven, e.g. ten consecutive videos needing no Gate-1 edits, not a vibe.
05 · The build
What we're building, and when
We evaluated three architectures (pure no-code n8n · full custom portal · thin portal over n8n) through three lenses: operator day-to-day, build load, and founder ROI. The middle path won. Because our AI stack is already running (n8n, Apify, fal.ai, kie.ai, Railway, and existing image-gen automations), the timeline is days, not weeks:
- Days 1-2: master template + proof video. Kyle designs the master trailer template and stands up the pipeline skeleton. Josh drives the creative testing: hero-shot recipe runs in fal/kie, voiceover auditions, music shortlist. Paul Mather's video is produced through the automation, exercising both gates for real. Hard checkpoint at 48h: if the template engine can't hit the "movie trailer" bar, we switch render engines before building further.
- Days 3-5: hardened automation + first delivery. The full pipeline runs end-to-end with asset archiving, completion webhooks, and missing-asset flagging live. Josh stress-tests both gates, builds the pronunciation dictionary (UK names like "Mitie" and "BAFE" trip the voiceover without it), and runs the review loop until Paul Mather's video is send-ready.
- Week 2: the review portal + first batch. Kyle ships the one-screen review portal (player, scene strip, fact-check panel, cost-labeled edit tiers, version history). Josh runs the first ~10 prospects through it end to end, owns the quality gate on every video, and tracks reply metrics so the founders see conversion data, not just output.
- Later, only if it works: upgrade the render engine for instant previews and ~$0.02 renders behind the same data contracts. A component swap, not a rebuild. We deliberately don't make that bet before reply data exists.
06 · Proof it matters
The first prospect, and the catch we already found
Paul Mather. Verified facts, safe to use:
- Managing Director since Oct 2017 (his LinkedIn headline says "Director of Business Development": a stale tagline, not his title)
- End-to-end UK fire & security: design, install, maintenance, 24/7 monitoring; NSI Gold and BAFE accredited
- Mitie completed its £350M acquisition of Marlowe plc in early August 2025 (~2,700 colleagues joining ~76,000)
The catch: the company has already rebranded. As of April 2026 contracts moved to Mitie Security Ltd trading as "Mitie Fire & Security", and marlowefireandsecurity.com now redirects to mitiefireandsecurity.com. A video proudly branded "Marlowe" would arrive looking out of date: the exact embarrassment this system is designed to prevent. Instead, the transition becomes our freshest personalization hook, and the pipeline screenshots the new site.
Sources conflict on employee count, so no video will ever cite a hard headcount. "Hundreds of engineers" is the verified-safe phrasing.
07 · The team
Roles
| Who | Role |
|---|---|
| Kyle | The brain and orchestrator: system architecture, pipeline build, AI prompt design, master template creative direction, campaign formula tuning |
| Josh | Execution owner: runs every video end to end (research review at Gate 1 → tweak loop at Gate 2 → delivery), owns the fact-check gate and the pronunciation dictionary, curates hero-shot takes and the music library, runs the deliverability test, tracks reply metrics per video |
| Luca / Neil | Value prop + CTA (§2), budget sign-off, campaign targets |
Josh's seat is the quality engine of this system: at 10-20 videos/month, roughly 30-60 minutes of his judgment per video is what separates "impressive personalized trailer" from "obvious AI spam", and his reply-metric tracking is how we'll know the channel works.
08 · The money
Costs, in one breath
Roughly $1-3 per video all-in (research, script, voiceover, AI shots, renders, re-rolls) and $200-300/month fixed at 10-20 videos, mostly cancellable monthly, so the experiment can be killed cheaply. It runs on accounts we already pay for (Railway, n8n, Apify, fal.ai, kie.ai) plus one standard Anthropic API key for the research AI.
09 · Eyes open
Risks we're actively managing
| Risk | Mitigation |
|---|---|
| Video looks "AI-generic," not cinematic | 48-hour aesthetic checkpoint before deeper build; AI footage limited to 1-2 atmospheric shots; no AI faces |
| Wrong facts about a prospect | Grounded research + source-tagged claims + automated verifier + human gates. Already caught the Marlowe→Mitie rebrand. |
| Wrong person (common names) | Name required at intake, hard-matched against the scraped profile; identity panel at Gate 1 |
| Music licensing | Stock licenses often exclude per-recipient generated video; pre-cleared library, checked once before volume sends |
| Email deliverability | Video links can hurt inbox placement; Josh runs a small deliverability test before scaling |
| LinkedIn scraping ToS | Public-data, cookieless scrapers only; no automated LinkedIn messaging, ever; GDPR basis + one-command data purge built in |
| Cost creep from re-rolls | Every asset cached forever; re-renders reuse, never regenerate; per-video and daily spend caps with a visible meter |
10 · Where we are
Status
- DoneDeep research on tooling & approach, independently fact-checked against live vendor pages, July 2026
- DonePaul Mather / Marlowe / Mitie facts verified with sources
- DoneArchitecture decided: three designs scored by independent review panels
- DoneBuild blueprints written & cross-checked: storyboard schema, AI prompts, asset pipeline, hero-shot recipes
- DoneProcess red-teamed end-to-end: ~40 failure modes with automated detection and recovery, folded into the operating spec
- BlockedValue prop + CTA from founders (§2)
- NextDays 1-2: master template + Paul Mather's video produced through the automation