EP 076 · 97 MIN STOP STARTING YOUR PROMPTS AT ZERO
How to Build AI Workflows That Never Start From Zero
The short version
Drew Brucker and Rory Flynn break down how they build AI systems that compound instead of resetting every time: ingesting call transcripts and emails as context, archiving proven sub-skills into a reusable vault, and running a 'prompt amplifier' skill before every request. They also cover LinkedIn's new, poorly defined 'AI slop' flag, Seedance 2.5's 30-second generations, Midjourney V8.2, and NVIDIA's new real-time synthetic-video detector.
Key takeaways
- 01
Drew's litmus test for a mature AI setup: your prompts should get shorter over time as skills and agents absorb the context you used to type out.
- 02
Feed recorded sales and client calls into a system daily; Drew pulls 'content seeds' and voice patterns from them instead of writing from a blank prompt.
- 03
Back in 2023, Rory mined 20 weekly sales-call transcripts with GPT-3.5 to find the most common objections, then wrote them into the sales script before they came up.
- 04
Using Fireflies sentiment data, Rory saw interest dip right before the pitch, so he moved price earlier into the call and closed more deals.
- 05
Rory archives proven sub-skills (used successfully about ten times) into a searchable 'vault' so an orchestrator agent can pull only tested building blocks.
- 06
Scrap skills as often as you build them: an unused tool just clutters the system, and stale ones drift out of sync with everything else.
- 07
NVIDIA's new detector reports 92% accuracy spotting synthetic video on uncompressed footage, built to run at the platform level instead of after-the-fact moderation.
Receipts · said on the record
“The real big issue: anybody can use this button. If I'm a business and I have a competitor, could I just go through and tag these things? I feel like there's a weaponization that could come into this.”
“We built whatever questions were asked the most into the sales script. So we were basically answering questions before they were asked.”
“Speed only matters in those circumstances, assuming you've got the right evals and context and things built into that.”
“It's really surface level because there's actually no … it's optimized for writing, not for thinking.”
“You have to think about the writing, but you have to think about the thinking. I have a whole kill list of AI LLM writing patterns that I weigh it against.”
Chapters
- 0:00:00 What is next for Fast Hours?
- 0:05:39 Why criticize AI while using it?
- 0:08:56 How can AI recover old creative work?
- 0:10:46 Why did 90s sports design feel better?
- 0:19:07 What is LinkedIn's AI slop button?
- 0:22:37 Why did LinkedIn reward lower effort?
- 0:29:10 How are creators gaming account reach?
- 0:32:43 Why do large platforms ship so slowly?
- 0:33:48 What changed in Seedance 2.5?
- 0:39:57 What is new in Midjourney V8.2?
- 0:43:48 How does texture improve AI prompting?
- 0:49:25 Why is Midjourney regaining momentum?
- 0:50:04 Why is AI entering its 3D moment?
- 0:58:12 Can people still detect synthetic video?
- 1:01:21 How does NVIDIA detect fake video?
- 1:07:10 Why does speed devalue AI creative work?
- 1:11:13 How do eval loops prevent AI slop?
- 1:13:00 How can calls create original content?
- 1:16:20 Why can't AI automate your opinion?
- 1:21:54 What should AI content systems ingest?
- 1:25:27 How can AI improve sales calls?
- 1:29:41 Why should prompts stop starting at zero?
- 1:31:08 How should AI skills be customized?
- 1:33:29 When should AI workflows be deleted?
- 1:35:12 How do you build an AI skill vault?
Questions this episode answers
What does Fast Hours mean by 'workflows that never start from zero'?
Drew Brucker and Rory Flynn describe a compounding AI setup where every prompt draws on saved context — recorded sales and client calls, archived proven skills, and a prompt-amplifier tool — instead of typing a fresh, unassisted prompt each time.
How did Rory Flynn use AI to improve his old sales calls?
Back in 2023, Rory fed roughly 20 weekly sales-call transcripts into ChatGPT 3.5 to surface the most common objections, then built those answers directly into his sales script and marketing. He later used Fireflies sentiment data to spot exactly when a call's energy dipped, so he moved his price discussion earlier in the call.
What did Fast Hours say about LinkedIn's new 'AI slop' button?
Drew Brucker criticizes it for applying only to posts (not the arguably worse comments), having no public definition of 'slop,' and being open to weaponization by competitors flagging each other with no clear consequence or review process. Rory ties it to LinkedIn's broader pattern of throttling reach for established creators while failing to reward new ones.
What's new in Seedance 2.5 and Midjourney V8.2?
Seedance 2.5 went live on Dreamina and BytePlus with native 30-second generations and support for up to 50 reference images. Midjourney's V8.2 is now out of preview, with V9 already in training and new edit models being tested on alpha.midjourney.com.
How does Rory Flynn organize his AI skills so he doesn't rebuild them?
He breaks larger pipelines into individual reusable sub-skills and archives only the proven ones (tested successfully around ten times) in a searchable vault, which his orchestrator agent can pull from for new projects. He also stresses scrapping unused or stale skills just as often as building new ones.
Show notes
A good AI workflow remembers what worked last time, so you don’t have to rebuild the same process from scratch.
Drew Brucker and Rory Flynn attempt to explain how advanced AI workflows actually work, then accidentally expose the mildly concerning machinery behind their own. The useful version of AI goes far beyond isolated prompts. It remembers prior decisions, pulls from real conversations, reuses approved skills, evaluates its own output, and saves the human from rebuilding the same process every Tuesday like a highly caffeinated goldfish.
The episode breaks down context files, AI agents, evaluation loops, prompt amplifiers, reusable skills, call transcripts, sales data, and personal knowledge systems. Drew and Rory also examine Seedance 2.5, Midjourney V8.2, emerging AI-assisted 3D workflows, synthetic video detection from NVIDIA, LinkedIn’s AI slop button, and the growing gap between generating faster and building systems that improve with use.
