6d • General discussion
RecapFlow : October 27th Coaching call analysis
📝 SUMMARY
This week's call packed an unusually wide range of member updates into one session. After some casual catch-up — including token-spend jokes and members comparing personal AI assistants like "Dot," "Hermes," and the newly-released "Duet" — Brandon Hancock led an extended infrastructure discussion on scaling parallel AI coding work-trees, covering compute, storage, and CI/CD cost bottlenecks (~$600/month in GitHub Actions) and trading ideas with Patrick Chouinard on NAS-based shared storage, multiple smaller Proxmox nodes, and hybrid NVMe-plus-spindle designs. The middle of the call turned into show-and-tell, with Patrick sharing an internal Claude "support skill" pipeline (97% opt-in in two weeks), a six-day autonomous "Fable" run, and a ~$6 AI avatar training video; Ty Wells, Paul Miller, Scott Rippey, and Juan Torres each demoed projects ranging from a Rust-based Adobe Fireworks clone to a "Stack Sentinel" monitoring app and a photo-booth AI image pipeline. Later segments covered Ryan C's rebuilt digital-signage business, Shakur Abdullah's CRM rebuild and FOIA/public-records workflow, a tip on reverse-engineering SaaS tools via GitHub, and a philosophical thread on AI-driven code review, before Brandon closed with an update that his company remains pre-revenue, awaiting acquisition or investment.
💡 KEY INSIGHTS
Patrick Chouinard reported extreme token usage polarization among go-live users.
He recommended separating storage on a shared NAS from compute on multiple smaller nodes rather than using one "monster" machine.
Patrick built a FinOps dashboard spanning Claude, Codex, OpenAI API, Wavespeed, Tavily, 11Labs, FireCrawl, and GitHub.
Patrick's agents use a base personality prompt nicknamed "Clara" and store API keys on a physical secret-manager server.
Patrick uses Sonnet 5.5 for heavy transcript ingestion and GLM 5.3 Flash for cheap bulk chunk-processing.
Brandon Hancock uses "Dot" as a co-pilot for managing GitHub issues, travel booking, and cancelling unused subscriptions.
Brandon suggested codifying every SDLC step into reusable skills or artifacts for leadership.
Shakur Abdullah built a "permit desk" in a single day, automating FOIA requests across Virginia counties.
Shakur Abdullah suggested AI adoption requires the "pain of not adopting" to be felt by the people being asked to change.
Patrick advised standardizing CI to only trigger at the PR layer, not every commit.
Scott Rippey built "Stack Sentinel" to scan Vercel, Netlify, and GitHub stacks daily for vulnerabilities.
❓ KEY Q&A
Q (Ty Wells): Is the video pipeline a GitHub repo or a product?
A (Patrick Chouinard): It's a GitHub repo — Scott Rippey's pipeline.
Q (Brandon Hancock): Why run both Dot and Hermes?
A (Patrick Chouinard): Hermes is a privileged infra agent (own email/calendar/GitHub) managing VMs/VLANs; Dot just manages his "online stuff" and watches GitHub.
Q (Brandon Hancock): How do you get around the parallel work-tree compute/storage bottleneck?
A (Patrick Chouinard): Use a good (ideally hybrid) NAS for shared fast storage and scale out with multiple smaller compute nodes rather than one big machine.
Q (Ty Wells): What's the name of that drive setup?
A (Patrick Chouinard): A "hybrid NAS" — NVMe cache in front of spindle drives.
Q (Brandon Hancock): VMs or different compute nodes?
A (Patrick Chouinard): A Proxmox cluster where he keeps adding nodes; with a good NAS, VMs can move live across nodes. For T3 code, install a T3 server on the VM so the local machine is just a client.
Q (Brandon Hancock): Is it one shared storage pool multiple nodes access, with work-trees on the NAS?
A (Patrick Chouinard): Yes; Synology for plug-and-play, or self-built TrueNAS-on-Proxmox for more advanced users, using NVMe cache in front of larger spindle drives.
Q (Paul Miller): Are you (Brandon) using hosted T3 server instances or coding on your Macs?
A (Brandon Hancock): Everything runs locally across his Mac Pro, Mac Mini, or new "Threadripper" machine, bought after his Macs maxed out from too many simultaneous work-trees.
Q (Brandon Hancock): Any "middle" models between cheap classification and flagship models?
A (Patrick Chouinard): Sonnet 5.5 for heavy transcript ingestion, GLM 5.3 Flash for cheap chunk processing; flagship models mainly for dev, not production. Hermes recently switched to "6.1 Tera," cheap and more capable than "Luna," good for ambient tasks.
Q (Juan Torres): Did you find a feasible media workflow with Wavespeed using the JSON file saved locally?
A (Patrick Chouinard): Yes, via Scott's pipeline; he mainly added inference providers, FinOps surveillance, and security.
Q (Brandon Hancock): Goal-and-run, or more steps to build the Fireworks/Sparkle clone in three days?
A (Paul Miller): A single high-level prompt asking Claude to deep-research Fireworks' look/feel/functionality and build a non-commercial equivalent for M-series Macs first.
Q (Brandon Hancock): Did you have a CI baseline before Blacksmith, and was this personal or company?
A (Paul Miller): Personal; heavy CI bottlenecks and repeated GitHub limit top-ups prompted the move, alongside hardware/network upgrades.
Q (Paul Miller, relaying his CTO's question): How do you get non-technical staff on board with a fast AI transformation?
A (Patrick Chouinard): Build hand-holding into Claude via a "support skill"/dispatcher that redirects normies to the right resource, paired with changelogs that auto-trigger regeneration of training material.
Q (Brandon Hancock): How does Dependabot compare to Stack Sentinel?
A (Scott Rippey): Dependabot is project-level (Next.js, React, etc.); Stack Sentinel also tracks hosting/runtime level — something Scott only recently learned while investigating the Node.js issue.
Q (Brandon Hancock): How does your tool compare to the other mentioned security tool ("CC Black Box"-type), and could you report back?
A (Scott Rippey): Hadn't used it, didn't know, but agreed to look into it and report back.
Q (Brandon Hancock): Does Tycode store any code or secrets, for HIPAA/SOC2 reasons?
A (Scott Rippey): Zero code and zero secrets stored — only results and fix statements — though he'd need to harden/document this further since it was built during beta.
Q (Brandon Hancock): What model made the cougar mascot image — was it a "scary face"?
A (Brandon Hancock): Image 2.5 "Flair" (from the new GPT 2.5 image model; "Sunburst" is the quality variant, but Flair is faster). It wasn't a scary face — he showed the real before/after.
Q (Brandon Hancock): Multiple reference images, or just one base image?
A (Juan Torres): Just one base image and a prompt; modern multi-modal diffusion models have enough reasoning to work from a basic prompt, unlike older stable-diffusion-era workflows.
Q (Patrick Chouinard): What models were used for the Claude Projects demo video, and how was it built?
A (Patrick Chouinard): All animation in Remotion (free); the only generative element was the speaking avatar via OmniHuman (chosen over Hagen/HeyGen at ~3x cost, per the transcript's varying spellings); audio via "Agent Turbo," ~1/10th ElevenLabs' price.
Q (Patrick Chouinard): Does the avatar's mouth move naturally despite Zoom's refresh rate?
A (Patrick Chouinard): Yes.
Q (Patrick Chouinard): How much did the whole generation (incl. A/B testing) cost?
A (Patrick Chouinard): About $6 total.
Q (Brandon Hancock): What avatar service did you use, and wasn't it ~3x more expensive ("Hadrian")?
A (Scott Rippey): Doesn't remember the specific model — Claude showed options and he picked one; the avatar was from a static generated picture, not a real photo of himself.
Q (Ty Wells): How is the remote screen system set up under the hood?
A (Ryan C): All CloudFlare Tunnel; the box installs an agent as device owner, and the dashboard communicates to the boxes through the tunnel.
Q (Scott Rippey): What Android device/model for the screens?
A (Ryan C): A commercial Android media player (~£150-160, Android 14, fully unlocked).
Q (Brandon Hancock): Go-to-market plan for the screens business?
A (Ryan C): SEO first, then Instagram/Facebook ads, possibly Google AdWords — SEO is already strong enough that he ranks on pages he'd otherwise pay for.
Q (Shakur Abdullah): How is it set up so a click actually triggers an action in T3 code?
A (Brandon Hancock / Juan Torres): A combination of webhooks; T3 code's new deployments handle job queuing, checking Remold and kicking processes back into T3 code.
Q (Juan Torres): Who are your testers trying to get FOIA information from?
A (Shakur Abdullah): Every county in Virginia, specifically building/permit-related requests.
🛠️ TOOLS AND CONCEPTS MENTIONED
Dot – OpenAI/ChatGPT-based personal assistant used widely for subscriptions, scheduling, GitHub issues, and as a coding co-pilot; several members gave theirs nicknames ("Jimothy," "Frank").
Hermes – Patrick Chouinard's privileged infra agent (own email/calendar/GitHub) managing VMs/VLANs.
Duet – new OpenAI feature/device demoed live; Ty Wells used it with Claude's design-system agent to attempt building a Toast POS replacement.
GrokBot – avoided by Paul Miller over trust/security concerns.
T3 / T3 code / T3Q code – coding platform/server; pending PR for ACP-based Hermes UI, planned "Mesh" multi-server feature, job-queuing/webhooks used by Shakur's Remold integration, MCP interface explored by Paul Miller.
ACP – protocol for T3-to-Hermes connectivity.
Claude Motion – teased tool ("Remotion on steroids"), runs on Opus 5.5, bundles ReMotion per Patrick.
Remotion – free underlying animation/rendering framework.
Wavespeed – media/diffusion-model aggregator ("OpenRouter for diffusion models"); primary supplier for Patrick, Juan Torres, and Scott Rippey; desktop app can build workflows via JSON.
Higgsfield – prior video supplier, swapped out for Wavespeed.
Open Router – multimedia model aggregator; Patrick avoids routing OpenAI models through it.
Proxmox – virtualization platform for Patrick's compute cluster/NAS/T3 server VMs.
Synology – entry-level plug-and-play NAS.
QNAP – more "enterprise" NAS option.
TrueNAS – self-built NAS option on Proxmox.
GitHub Actions – CI/CD service, costly at scale (~$600/month for Brandon).
Blacksmith – CI alternative adopted by Paul Miller.
Depot – CI alternative suggested by Shakur Abdullah in chat.
FinOps dashboard – Patrick's custom spend-tracking tool across Claude, Codex, OpenAI API, Wavespeed, Tavily, 11Labs, FireCrawl, GitHub.
Tavily, 11Labs/ElevenLabs, FireCrawl – API suppliers in Patrick's pipeline/dashboard.
GLM 5.3 (Flash) – cheap model for bulk/ambient processing.
Sonnet 5.5 – model for heavy transcript ingestion.
Opus (5.5 / 6.1 high) – flagship dev models; also Vercel AI SDK agent basis for Symbiote.
Claude "6.1 Tera" – model Hermes switched to.
"Luna" / "deep seek flash" – cheap classification-tier models.
Potato – community figure/approach referenced for PR-minimization strategy.
UltraCode – platform for "Fable" runs and Scott's gym-system build.
Agent Ops – Patrick's network-management platform used by Hermes.
OLAMA server – local model server substituted for Databricks/Intra-ID equivalents.
GitHub Copilot – coding assistant referenced for memory-system compatibility.
ChatGPT – now with voice + read/write GitHub access for issue creation.
Rust – language used to rebuild Adobe Fireworks ("Sparkle"/"Sparkler").
dot ploy – deploy tooling for Paul Miller's VPS servers.
Vercel AI SDK – used to build "Symbiote."
Symbiote – Scott Rippey's CrossFit/Hyrox gym programming AI with coach-feedback loop.
Stack Sentinel – Scott Rippey's stack-vulnerability monitor (Vercel/Netlify/GitHub, daily alerts, embedded chat).
Vercel / Netlify – hosting platforms criticized for not updating Node.js runtimes.
Tycode (chat log shows site tightcode.dev, promo "brandon20") – Scott Rippey's code-review product with untested HIPAA/SOC2 agents; stores no code/secrets.
Dependabot – basic project-level dependency scanner, compared against Stack Sentinel.
Model Radar – Scott's earlier beta Mac app, being folded into Stack Sentinel.
CC Black Box (ccblackbox.app) – free Claude Code companion app (Mac + iOS), incorporating some T3 features into its agent harness per Scott Rippey.
WireGuard / Tailscale / TeamViewer – networking/remote-access options; Juan Torres uses WireGuard/Tailscale-style secure connection; Ryan C rejected Tailscale/TeamViewer at scale in favor of CloudFlare Tunnel.
AWS EC2 / CloudFront – Juan Torres's photo-booth backend infrastructure.
Image 2.5 (Flare/Flair / Sunburst variants) – GPT's image generation model for mascot-style transforms.
OmniHuman – avatar animation model (chosen over HeyGen).
HeyGen ("Hagen") – alternative avatar tool, ~3x OmniHuman's cost.
Agent Turbo model – audio/voice generation, ~1/10th ElevenLabs' cost.
FailPoint AI – competing video supplier evaluated by Juan Torres.
Hugging Face / Civic AI – open-source model repos combinable with proprietary models in Wavespeed workflows.
"Hadrian" / "Heijan" / "H&M" – ambiguous name(s) for an avatar/face-generation service discussed by the group.
CloudFlare Tunnel – backbone for Ryan C's remote screen-control system.
ADB (Android Debug Bridge) – used to connect to Android media players.
Nano Banana – image model powering Ryan C's storefront visualizer.
Instantly – cold email outreach tool (needs ~2-week warm-up), recommended by Brandon Hancock.
Resend – email delivery backend for Ryan C's campaign tool.
Claude Design – design tool now drivable directly from Claude Code.
Slop Monster – GitHub writing tool to make AI copy read less like AI content.
EMS Soap – Ryan C's cold-outreach scraping project.
Remold – Shakur Abdullah's tool being redesigned for T3 code compatibility.
ngrok – raised as a possible (unconfirmed) local-to-cloud tunneling option.
"Dots" – referenced as having "jumped on" and handled some automated FOIA email replies.
GitHub "reverse engineer" repo (morluto/rea) – suggested to reverse-engineer existing tools like MailChimp.
MailChimp – example target for the reverse-engineer approach.
CLIProxyAPI (router-for-me) – suggested by Patrick to Brandon for managing multiple subscriptions.
Infisical – secret/API-key management, free self-hosted or paid hosted, part of Patrick's security stack.
step-ca (smallstep) – certificate authority, part of Patrick's security stack.
Traefik – reverse proxy, part of Patrick's security stack.
Uptime Kuma – uptime monitoring, part of Patrick's security stack.
CorridorKey 2.0 – greenscreen tool Ryan C suggested Patrick integrate into his video pipeline.
Minisforum AI X1 Pro (Walmart listing, "Ryzen AI 9 HX-370") – Brandon Hancock's new local CI/compute machine.
Asustor NAS – suggested NAS product (Andrew Nanton).
treg.to – self-hosted tool Andrew Nanton asked about; unclear to the group what it does.
UI.com Cloud Gateway Fiber (UCG Fiber) – router recommended by Patrick to Paul Miller for his 2Gbps connection.
Plex / JustWatch API – Alex Wilson's movie-recommendation app ties into Netflix history via JustWatch API and outputs to a Plex list.
Theo's skills collection ("tstack") – shared by Patrick Chouinard; Shakur Abdullah reviewed it positively.
📎 SHARED RESOURCES
Links and resources shared during the call:
https://littlefarmerscay.com (Ty Wells' family eco-tourism venture, referred to in-call as "Little Farmers Key")
🔄 FOLLOW-UPS WORTH EXPLORING
Here are the open questions, follow-ups and commitments from this week's call worth revisiting:
Brandon Hancock will pursue a potential second Claude ($500) subscription next week to maximize token usage.
Brandon Hancock will investigate "Potato"-style smaller/quicker PR techniques and report back to the group; he has also adopted draft-PR workflows to cut CI runs.
Patrick Chouinard is building a standardized CI workflow so CI only runs at the PR layer, not every commit.
Patrick Chouinard will keep his fork of Scott Rippey's video pipeline private until secrets are secured, then make it public.
Patrick Chouinard will share his FinOps dashboard with the group once sensitive financial details are blurred.
Paul Miller will share his Blacksmith CI setup/results with Brandon Hancock (he moved his CI from GitHub Actions to Blacksmith the morning of the call).
Paul Miller will continue building an MCP-based wrapper around T3 code for a video-brief generation pipeline.
Ty Wells will follow up with Patrick separately on infrastructure after testing his own setup.
Ty Wells will keep the group posted on the Little Farmers Key/Cay Bahamas venture.
Ty Wells will look into adopting "Instantly" for cold outreach.
Scott Rippey will consider turning Stack Sentinel into a SaaS product if there's enough interest, though he is undecided on liability/security concerns.
Scott Rippey will give Brandon Hancock (or his contact) free access to Tycode to test the untested HIPAA/SOC2 agents.
Brandon Hancock will review his organization's vendor compliance requirements before integrating a tool like Tycode.
Scott Rippey will follow up offline with Brandon Hancock on Tycode's compliance documentation, and will look into the comparable security tool Brandon mentioned and report back.
Juan Torres will post a more technical video about his event infrastructure to LinkedIn, and will share his first "We Hope It's Good" podcast episode once recorded.
Juan Torres will move forward with WaveSpeed as his chosen video generation platform.
Ryan C will pause further coding this month (after a coffee spill on his MacBook) to focus on sales/video/photo for screen clients, while continuing to build outreach/scraping features into his CRM.
Patrick Chouinard will find and share the "Slop Monster" GitHub repo link (subsequently posted by Ryan C in chat).
Shakur Abdullah will continue rebuilding his CRM around "T3 code" and developing "permit desk," including eventually connecting it to an automated mailing service; he will send Remold to testers once redesigned for T3 code compatibility.
Ryan C will try pointing a GitHub "reverse engineer" repo at MailChimp once his usage credits reset; Ty Wells may try the same.
Ryan C will review FOIA automation guidelines to confirm county-specific bureaucratic compliance.
Brandon Hancock will prepare and share a more fleshed-out plan of his project's direction in a future recorded session, possibly including Ty Wells.
0
0 comments
RecapFlow : October 27th Coaching call analysis
skool.com/ai-developer-accelerator
Master AI & software development to build apps and unlock new income streams. Transform ideas into profits. 💡➕🤖➕👨💻🟰💰
Leaderboard (30-day)
1
+1
2
+1
3
+1
4
+1
5
+1
Powered by