Prompt and Circumstance

We show you how to run YuE2, an open-weights music model, locally in ComfyUI. Suno is having legal troubles and quality is getting worse, so you need a model that gives high performance and without paying for a subscription. You'll learn how YuE2 gives high quality music by generating sheet music first, how to set it up in Comfy UI with some time-saving tricks, and you'll hear some examples of how it handles different types of music.

David is an expect in local media generation, who has tested hundreds of open weight models. He shows you the easy setup, the tricks to save you time, and the best models to use so you don't waste your time. All you need for this week is 4GB of VRAM, but 8+GB will get you faster performance on a higher-bit quantization of YuE2.

Don't miss this opportunity to learn how to generate media locally.

Chapters

00:00 - Why Users Are Leaving Suno
00:57 - What Makes YuE2 Different
02:11 - Setting Up YuE2 in ComfyUI
04:42 - Generating and Hearing the Results
06:48 - Weaknesses and Lo-Fi Use Cases
07:44 - License Limits for Commercial Use

Links

YuE2 on Hugging Face - https://huggingface.co/m-a-p/YuE2-3B
YuE2 on ComfyUI Wiki - https://comfyui-wiki.com/en/models/yue2
ComfyUI - https://www.comfy.org
Suno - https://suno.com
Devil Wears Product shop - https://devilwearsproduct.shop

Find Us

YouTube - https://www.youtube.com/@PandCpodcast
Bluesky - https://bsky.app/profile/pandcpodcast.bsky.social
X - https://x.com/_pandcpodcast
Instagram - https://www.instagram.com/_pandcpodcast
LinkedIn - https://www.linkedin.com/company/p-and-c-podcast

Jev is the new "System 1" AI model from TypeSafe AI, and it doesn't write you an essay. It just answers. We break down what Jev is, why it's trained differently from ChatGPT or Claude, and two ways to use it today, no developer skills needed. The full setup prompt from the episode is below, so you can copy it.

Jev is trending on X right now. It's a model that returns a typed, calibrated decision with a confidence score instead of a wall of text. TypeSafe claims it's about 194x faster and 445x cheaper than a general-purpose LLM on the tasks it's built for. We put that to the test with two real examples, classification and semantic search, head to head against Sonnet on cost and speed.

In this episode:

→ System 1 vs System 2 thinking (h/t Kahneman's Thinking, Fast and Slow)

→ RLHF vs RLCD: why Jev doesn't get sycophantic

→ How to get API access and set Jev up as a skill in your AI agent

→ Real numbers: Jev vs Sonnet on cost and latency

→ How Jev stacks up against GPT-5.6 Luna on classification

→ Jev's three output types, and why every answer comes with a confidence score

[00:00] Introduction

[00:25] What is Jev

[01:18] RLHF vs RLCD explained

[02:22] How to set up Jev (API + agent skill)

[03:44] Example 1: Classification vs Sonnet (cost)

[04:46] Jev vs GPT-5.6 Luna

[05:05] Example 2: Semantic search vs Sonnet (speed)

[06:25] Output types and confidence scores

[06:51] Final thoughts

━━━━━━━━━━━━━━━━━━━━

THE JEV SETUP PROMPT (copy and paste this into Claude, or into your AI agent)

━━━━━━━━━━━━━━━━━━━━

We're going to add an AI model called Jev to this workspace.

Jev is made for sorting and deciding. It generates typed responses, not text. It only has 3 types of responses: pick from an option on a list, pick a score from a scale, or say how likely something is true. It also usually tells you how sure it is.

Some notes:

* Jev is a new type of model that is not in your training set. Please read how it works before proceeding: https://docs.typesafe.ai/llms.txt

* I'm going to use it through typesafe directly. Here is the prompt for how to set it up: https://docs.typesafe.ai/agent-skill

* ask me for my TypeSafe key (I got one from https://typesafe.ai) and keep it in my .env file

After we add Jev, perform a test. Create 10 feature requests for the same product and score them on impact and effort. Show a side by side comparison against scoring those same requests in Claude using Sonnet in a way that I can see the speed and the cost of classification with Jev vs Sonnet.

Also ensure that the skill you add has an off switch that you can explain to me how to use so that I can enable or disable Jev as I wish.

Anything you send to Jev leaves my computer, so ask me before sending anything private.

From now on when Jev is enabled, you let Jev make decisions while you write. If Jev isn't sure then you take over.

━━━━━━━━━━━━━━━━━━━━

Sources for the numbers and screenshots in this episode:

• TypeSafe AI, "Introducing System One Models & Jev": https://typesafe.ai/blog/introducing-system-one-models-and-jev

• MarkTechPost, "TypeSafe AI Releases Jev": https://www.marktechpost.com/2026/09/19/typesafe-ai-releases-jev/

Subscribe to Prompt & Circumstance for practical AI breakdowns you can use today.

FIND US

YouTube - https://www.youtube.com/@PandCpodcast

Bluesky - https://bsky.app/profile/pandcpodcast.bsky.social

X - https://x.com/_pandcpodcast

Instagram - https://www.instagram.com/_pandcpodcast

LinkedIn - https://www.linkedin.com/company/p-and-c-podcast

Learn the full tool stack behind a production vibe-coded app: Claude Code, Cowork, Notion, Linear, Vercel, and Supabase. Phil Mau walks David and Ilan through building Shikomi, a recipe app he built solo with Claude Code, and explains why the real breakthrough wasn't the tools but treating AI like a team instead of a chat window.

CHAPTERS

00:00 - Why So Few Vibe-Coded Apps Ship
00:55 - Meet Phil Mau: Product Manager Turned Builder
01:32 - The Job Search That Sparked an AI Pivot
02:43 - From Claude Chat Chaos to a Notion and Linear Workflow
07:05 - The Mindset Shift: Treat AI Like a Team
08:23 - Demo: Shikomi's Cook Mode, Version History, and Where to Find Phil

SPONSORS

Devil Wears Product → https://devilwearsproduct.shop

LINKS

Shikomi - https://getshikomi.com/
Phil Mau (LinkedIn) - https://www.linkedin.com/in/mauphilip/
Phil Mau (Website) - https://philmau.com/

FIND US

YouTube - https://www.youtube.com/@PandCpodcast
Bluesky - https://bsky.app/profile/pandcpodcast.bsky.social
X - https://x.com/_pandcpodcast
Instagram - https://www.instagram.com/_pandcpodcast
LinkedIn - https://www.linkedin.com/company/p-and-c-podcast

Faceless explainer videos are a great way to engage your audience and with AI you can generate them in just a few minutes.

We teach you how to turn a quick audio recording into a faceless explainer video with AI generated images, starting from reference screenshots and ending with a finished clip stitched together. The whole tutorial takes 6 minutes.

CHAPTERS

00:00 - Cold Open
00:32 - What Faceless Explainer Videos Are
01:53 - Loading Reference Images Into ChatGPT
02:44 - Testing the AI Generated Style Prompt
03:27 - Generating a Full Video From a Script
05:06 - Turning the Process Into a Repeatable Skill

SPONSORS

Devil Wears Product - devilwearsproduct.com

LINKS

ChatGPT - https://chatgpt.com
Gemini - https://gemini.google.com
Claude - https://claude.ai
ffmpeg-python (GitHub) - https://github.com/kkroening/ffmpeg-python

FIND US

YouTube - https://www.youtube.com/@PandCpodcast
Bluesky - https://bsky.app/profile/pandcpodcast.bsky.social
X - https://x.com/_pandcpodcast
Instagram - https://www.instagram.com/_pandcpodcast
LinkedIn - https://www.linkedin.com/company/p-and-c-podcast

How to make a first-frame-to-last-frame video with LTX 2.3 in ComfyUI using just two images and a text prompt, running locally on consumer hardware.

David walks through the exact ComfyUI workflow behind our last episode's intro: the FP8 quantized version of LTX 2.3 (so you don't need an enormous GPU), the FLF2V template, and how a single prompt tries to bridge your first and last frame into a 5-second clip. We also show what happens when the first generation doesn't listen to the prompt, why the noise seed matters, and how re-rolling, plus a Claude-refined prompt, changes the result.

Important: LTX-2.3 video generation requires CUDA 13+ and an Nvidia GPU

What's covered:
• Where to find the FLF2V (first-frame/last-frame-to-video) template in ComfyUI
• Downloading the right quantized LTX 2.3 + Gemma 3 models
• Setting dimensions, FPS, and clip length
• Writing and troubleshooting a video prompt
• Why your seed affects the output more than you'd think
• Using an LLM (Claude) to rewrite a weak prompt into a better one

Chapters:
0:00 Intro
0:14 Building the Workflow in ComfyUI
2:18 Setting the Prompt & First Generation
3:06 Reviewing the First Result
3:49 Troubleshooting: Seed & Reroll
4:09 Second Attempt, Better Prompt
5:09 Wrap-up

Merch:
https://devilwearsproduct.shop

Links:
https://ltx.io/blog/run-video-generation-model-locally

https://github.com/Lightricks/ComfyUI-LTXVideo

https://ltx.io/model/ltx-2-3

https://ltx.studio/ 

Click here to watch a video of this episode.

#LTX2 #ComfyUI #AIVideo

Most AI image models fight you when you want flaws. Krea 2 doesn't. We dig into why this open-source, locally-run model is so good at "imperfect" realism (fisheye distortion, blur, grain), then walk through setting it up in ComfyUI end to end: grabbing the template, downloading the missing models, and picking resolution by aspect ratio instead of raw pixels. It works well with text and the template workflow has built-in LoRA support!

Chapters
00:00 - Krea 2 Nails "Imperfect," Realistic Images
00:51 - What Krea 2 Is
01:15 - Setting Up the Krea 2 Template in ComfyUI
02:25 - Workflow Options: Prompt Enhancement & Resolution
03:42 - First Generation: Speed Test at 1 Megapixel
04:24 - Scaling Up to 4K
04:56 - Testing Text Rendering on a Whiteboard
05:30 - Adding LoRAs for Style: The Rainy Window Example
07:01 - Wrap-Up

Merch
https://devilwearsproduct.shop

Links
Krea 2 - https://www.krea.ai/krea-2
Krea 2 Open Source - https://www.krea.ai/krea-2-open-source
Krea AI LinkedIn - https://www.linkedin.com/company/krea-ai/
Krea 2 Licensing - https://www.krea.ai/krea-2-licensing
Krea 2 Technical Report - https://www.krea.ai/blog/krea-2-technical-report

Click here to watch a video of this episode.

OpenClaw can move product monitoring from simple error alerts toward self-healing agent workflows. Ilan explains the progression and patterns to use to go from “agent notices an error and tells you,” to “agent diagnoses and recommends a fix,” to a guarded autonomous flow where agents clone, patch, test, and eventually promote fixes into production. The big idea: close the loop only as far as you trust the system, and optimize for the customer never feeling the bug.

Chapters
00:00 - Why Closing the Loop Matters for AI Agents
03:00 - The Recommended Fix Flow: AI Diagnoses, Human Implements
04:00 - The Current State: Two-Model Monitoring with Human Approval
07:02 - The Fully Autonomous Flow: Clone, Test, Deploy
07:56 - Proof of Fix and What Comes Next

Links
OpenClaw - https://openclaw.ai
n8n - https://n8n.io
Click here to watch a video of this episode.

Find Us
YouTube - https://www.youtube.com/@PandCpodcast
Bluesky - https://bsky.app/profile/pandcpodcast.bsky.social
X - https://x.com/_pandcpodcast
Instagram - https://www.instagram.com/_pandcpodcast
LinkedIn - https://www.linkedin.com/company/p-and-c-podcast

Google released Gemini Omni at Google I/O 2026 and we ran it through its paces in Google Flow to see if the hype holds up. It mostly doesn't. At least not yet.

Gemini Omni is Google's new multimodal AI model, meaning one model handles text, images, audio, and video. You access it through Gemini directly or through Google Flow, and a paid Google subscription is needed. The flagship use case right now is AI video generation, so that's what we came to test. We tried uploading source footage to remix and style-transfer. The uploads stalled at 0% and never moved. Different upload path. Same result.
So we switched to image-to-video: upload a photo, write a prompt, get a clip. The upload worked. The prompt didn't. Gemini Omni generated something, but it ignored the source image and went cartoonish in a direction we never asked for. Adjusting the prompt made things worse.

Subscribe to Prompt and Circumstance for hands-on tests of what's actually shipping in AI.

[00:00] - Intro to Google Omni
[04:04] - Uploading Video
[06:38] - Uploading an Image
[10:57] - Conclusion

LINKS
Google Omni: https://deepmind.google/models/gemini-omni/
Contrast this to something more reliable: https://youtu.be/Y-Saa82qPX0
Media links:
https://pixabay.com/videos/mid-autumn-stone-staffordshire-90470/
https://pixabay.com/videos/hiking-walking-trail-man-alone-282995/
https://pixabay.com/photos/girl-child-teenager-dog-pet-small-5087960/
Video: Click here to watch a video of this episode.

#GeminiOmni #GoogleOmni #AIVideo #GoogleFlow #GoogleIO2026 

David walks Ilan through a local ComfyUI workflow that chains three AI models to upscale generated images from 720p all the way to 4K, with a side mission of fixing that plastic-y AI skin look.

Chapters

00:00 - Intro and what we are upscaling today
00:36 - Workflow overview and chaining three models
01:30 - Installing the ComfyUI Manager
02:30 - The download note and where models go
03:00 - Walking through CivitAI to grab the first 22GB model
05:00 - Epic Realism and SeedVR2 from ByteDance
06:00 - LoRAs, CLIP, VAEs, and refreshing node paths
08:00 - Generating the test image and loading the upscale workflow
09:30 - Running the three-pass upscale and what each model does
10:30 - Before and after at 4500 by 2500
12:00 - Final image, runtime, and going to 8K

Hilarious Product Management Merch
https://devilwearsproduct.shop

Links
Workflow: https://pastebin.com/SUtVN2Hh 

Password: Prompt and Circumstance Podcast

Diffusion Models (models/diffusion_models):

UltraFineTune: https://civitai.com/models/978314/ultrareal-fine-tune

ZEpicrealismturbo FP8: https://civitai.com/models/2305301/z-epicrealism

Checkpoints (models/checkpoints):

Epic Realism XL Crystal Clear: https://civitai.com/models/277058/epicrealism-xl?modelVersionId=1920523

Loras (models/loras)

Aidma Realistic Skin: https://civitai.com/models/1157318/photorealistic-skin-no-plastic-flux

SDXL Skin Realism: https://civitai.com/models/248951/skin-realism-acne-skin-details-imperfections-sdxl

CLIP

Clip_l: (models/clip) https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors

Tt5xxl fp8: (models/text_encoders) https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp8_e4m3fn.safetensors

Text Encoders:

Qwen 3 4b: (models/text_encoders) 

https://huggingface.co/Comfy-Org/z_image_turbo/blob/main/split_files/text_encoders/qwen_3_4b.safetensors

Variational Autoencoders (models/upscale_models)

Flux VAE (ae.safetensors): https://huggingface.co/lovis93/testllm/blob/ed9cf1af7465cebca4649157f118e331cf2a084f/ae.safetensors

Upscale Models (models/upscale_models)

4x Clear Reality: https://huggingface.co/skbhadra/ClearRealityV1/commit/bc01e27b38eec683dc6e3161dd56069c78e015ac

SeedVR2 Models (models/SEEDVR2)

Model: https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/seedvr2_ema_3b_fp8_e4m3fn.safetensors

VAE: https://huggingface.co/numz/SeedVR2_comfyUI/blob/main/ema_vae_fp16.safetensors

Find Us
YouTube - https://www.youtube.com/@PandCpodcast
Bluesky - https://bsky.app/profile/pandcpodcast.bsky.social
X - https://x.com/_pandcpodcast
Instagram - https://www.instagram.com/_pandcpodcast
LinkedIn - https://www.linkedin.com/company/p-and-c-podcast

Set up a decentralized agent team in OpenClaw to complete complex tasks. We walk you through a real use case to set up a morning briefing using two agents, a researcher and a summarizer. They hand off work to each other on a schedule so you wake up to a one-minute AI news brief on your phone.
 
Chapters
 
00:00 - Intro
00:47 - Why You Need an Agent Team
02:50 - Agent Team Patterns
04:50 - Ilan's Morning Briefing Setup
08:30 - How Zoe Was Built
10:50 - Wrap-Up and Key Takeaways
 
 
Business Use Cases
 
- Set up a decentralized agent team in OpenClaw where each agent saves its output to a workspace file and the next agent picks it up to complete its part of the task.
- Use cron jobs to schedule agents in sequence with time gaps between them, then test each agent manually before automating.
- Give each agent a pop culture persona that matches the work style you want, and define a clear output contract (a saved file) so the agent knows exactly when its job is done.
 
 
Links
 
OpenClaw - https://openclaw.ai
OpenClaw GitHub - https://github.com/openclaw/openclaw
OpenClaw Docs - https://docs.openclaw.ai
Agent Team Builder Skill - https://github.com/canuckamok/agents/tree/main/skills 
 
Find Us
 
YouTube - https://www.youtube.com/@PandCpodcast
Bluesky - https://bsky.app/profile/pandcpodcast.bsky.social
X - https://x.com/_pandcpodcast
Instagram - https://www.instagram.com/_pandcpodcast
LinkedIn - https://www.linkedin.com/company/p-and-c-podcast

• (00:00) - Intro