The payoff is a concrete pipeline: transcribe every clip with Whisper, let the AI find the story in the words, review the proposed script on a simple local page, then chop the footage into handled clips ready to drop into the editor's own Final Cut Pro project. It took a two-day job down to about an hour — not by finishing the edit, but by killing the tedious hunt through hours of footage.

In this episode

• The ambitious approach — AI watches footage and cuts its own version — failed on timing, and piling on B-roll and lower thirds only made a shaky thing worse.
• The trick was to stop asking AI to look and start asking it to read: transcripts turn footage into timestamped words, which is what these models are actually good at.
• The pipeline outputs sorted clips with handles (extra room front and back), not a finished video — the taste and the final cuts stay with the human.
• A one-to-two-day job dropped to about an hour, mostly because nobody had to scrub hours of footage hunting for usable moments.
• It's cheaper than you'd expect: the heavy lifting happens in free command-line tools, so the AI is only doing text work — doable on a $20/month plan.
• The honest split: talking-head footage (interviews, testimonials, webinars) works well because it's text; purely visual asks work but are less reliable, which matters when a client is watching.

Timestamps

• 0:00 — The question and the two-day-to-one-hour teaser
• 0:49 — Editing vs generating, and keeping your own process
• 1:55 — What failed: AI watching and cutting footage
• 2:38 — The fix: read instead of look, with Whisper transcripts
• 3:06 — Reviewing the script, then exporting clips to Final Cut
• 3:38 — Before and after: two days down to an hour
• 4:30 — What actually comes out: clips, not a finished video
• 4:49 — What it takes: tools and surprisingly low cost
• 5:45 — Who can do it: terminal work required
• 6:00 — What to watch for: words vs pictures, and buying time
• 7:18 — Send in your question

Tools mentioned: Whisper, Final Cut Pro, FFmpeg, Claude Code, Codex.

Got a question you've been wondering about — can AI do this thing? Send it to drew@workandwhistle.co.

AI Honestly is a Work and Whistle production. More at workandwhistle.co.

The real substance is the question underneath: buy off-the-shelf or build custom? Drew walks through exactly how he's built these systems — a VPS running Claude Code, Postmark to catch and send email, trained on your site and a page of rules — and argues that because the big platforms only sell 'speed to lead' bundled into full nurture suites, most people are pushed toward custom. He closes with the things the salesperson won't tell you: where the human takes over, whether to disclose the AI, and the buy-versus-build tax.

In this episode

• The MIT-cited stat: responding to a lead within five minutes instead of thirty makes you roughly 100x more likely to reach them — while the average business takes over a day.
• People send inquiries to three or four businesses at once, so the first one to reply with something that isn't garbage usually wins the deal.
• Under the hood it's simpler than people think: a VPS with Claude Code (or just a serverless function), Postmark to receive and send email as structured data, and an agent trained on your site, docs, or a page of rules.
• Real response times run about four or five seconds — Drew undersells it as 'under a minute' on his own site.
• Off-the-shelf options like HubSpot and Salesforce bury lead-answering inside a whole nurture suite, so you buy the house to use the front door — which is why Drew says it mostly points to custom, a couple-days-to-a-week build.
• Three things the seller won't mention: decide where the human takes over, choose whether to disclose it's AI (getting caught pretending is worse than saying so), and know you're paying either a platform tax or an upkeep tax.

Timestamps

• 0:00 — The question: AI that answers leads instantly
• 0:30 — Show intro
• 0:43 — What 'speed to lead' really means
• 1:09 — Why speed wins: the MIT stat
• 2:05 — How it works under the hood
• 2:40 — The build: VPS, Postmark, Claude Code
• 3:24 — Custom vs. off-the-shelf
• 4:04 — What a custom build takes
• 5:08 — Platform users and Request Router
• 5:48 — What to watch for: handoff, disclosure, cost
• 6:52 — Wrap-up and send your question

Tools mentioned: Claude Code, Postmark, Request Router, HubSpot, Salesforce.

Got a question you've been wondering about — can AI do this thing? Send it to drew@workandwhistle.co.

AI Honestly is a Work and Whistle production. More at workandwhistle.co.

In this episode

• The full workflow: talk into a voice memo, turn the auto-transcript into a script with AI, clean out the AI 'tells' by hand, then render it through an ElevenLabs voice clone.
• AI gets a script about 80% of the way; the last 20% is Drew manually cutting the em-dashes and 'it's not this, it's that' phrasing people can smell.
• A counterintuitive voice-cloning lesson: a too-polished clone sounds fake, and Drew's naturally flat, monotone delivery actually hides the AI tells.
• Real costs are small but watch the credits — ElevenLabs Creator plan is $22/month and re-rendering edits burns credits fast; Cloudflare hosting runs to pennies.
• A Claude Code pipeline publishes everything on command: show notes, the feed Apple/Spotify/Amazon pick up, a website page, analytics, and auto-extracted social clips.
• The honest catch: AI solved production, not distribution — getting anyone to actually listen is still the real, unsolved job.

Timestamps

• 0:00 — The question: can AI make a whole podcast?
• 0:13 — What AI Honestly is
• 0:29 — Why people ask: skipping the tedious parts
• 0:57 — The workflow: just talk into a phone
• 1:13 — Turning a transcript into a script
• 1:33 — Cutting the AI tells by hand
• 2:01 — ElevenLabs, voice cloning, and music
• 2:20 — Why polished isn't always more real
• 2:47 — What it costs to do it yourself
• 3:13 — Publishing with Claude Code and Cloudflare
• 4:02 — Analytics and auto-generated clips
• 4:24 — The catches: pipeline setup time
• 4:45 — AI fixed production, not being heard
• 5:05 — So, can you do it? Yes, with a catch
• 5:30 — Send in your question

Tools mentioned: ElevenLabs, Claude Code, Cloudflare.

Got a question you've been wondering about — can AI do this thing? Send it to drew@workandwhistle.co.

AI Honestly is a Work and Whistle production. More at workandwhistle.co.

Every month that email platform bill shows up, and it climbs as your list grows — because you're basically paying to store email addresses. So can you just rebuild the whole thing with AI and pay way less? Short answer: mostly yes, but "cheaper" comes with a trade. This week I break down the three versions of that answer, walk through the exact setup I've used, and flag the stuff the savings pitch leaves out.

What's covered:

• Why email platforms charge more as your list grows, and what you're actually paying them for (hint: it's deliverability, not the dashboard)
• The three versions of "replace it with AI" — the basics, the custom bonus layer, and the top-end rebuild that's probably not worth it
• The setup: Resend for sending, Supabase for storing your list, and AI agents to run the whole thing
• Real costs — Resend's free tier (100 emails/day, 3,000/month) and the $20/month paid tier
• How to build in unsubscribes the right way with a Supabase edge function
• A custom example: auto-sending a one-time thank-you email on a customer's first purchase
• Putting a simple interface on top so your team can send without touching a command line
• Three things to watch for: owning your own edge cases, rebuilding advanced automations, and protecting your sender reputation on shared tiers

Tools mentioned: Resend, Supabase.

Got a question you've been wondering about — can AI do this thing? Send it to drew@workandwhistle.co.

AI Honestly is a Work and Whistle production. More at workandwhistle.co.