2d • General discussion
RecapFlow : September 29nd Coaching call analysis
📝 SUMMARY
This was a long, wide-ranging call that opened with OpenAI Dev Day excitement (GPT-6.1 Sol, the "Dots" ambient agent, GPT Space, a Decision API, and a $500/month Cerebras tier) and closed hours later with casual banter. In between, Brandon Hancock demoed his agentic software-development workflow and shared a startup update, while other members — Daniel Zivkovic, Shakur Abdullah, Andrew Nanton, Tom Welsh, and Patrick Chouinard — showed parallel workflows and tools (Fable, Herder, T3 Code, CMUX) and business updates, including Tom's Asset MS product sold to farmers and Ty Wells' rapidly scaling ERP business. A major throughline was the OpenAI-vs-Anthropic competitive debate, with Patrick predicting Anthropic's response to Dev Day and the group debating whether Anthropic's enterprise-first posture will cost it long-term. Scott Rippey demoed CC Blackbox and Tight Code plus his AI video pipeline, drawing go-to-market advice from Brandon. The back half shifted into hands-on coaching: Patrick's home-lab "Agent Operated Environment," Brandon's cold-outreach and data-pipeline tactics, and heavy group coaching for Juan Torres on his AI photo booth business — pricing, seasonal launches, and sales targeting. Paul Miller added SaaS-founder perspective on VC dynamics and rapid prototyping, and Brandon closed with candid reflections on financial strain and gratitude for the group.
💡 KEY INSIGHTS
Model releases and strategy
OpenAI Dev Day shipped GPT-6.1 Sol (same price as 5.6 Sol, ~4x cheaper than Astra/Fable-class), "Dots" (a proactive ambient agent across chat platforms and phone, paired with a cheap/fast Decision API), GPT Space (a ChatGPT-native Microsoft 365 competitor), and a $500/month ultra-fast Cerebras tier at ~3,000 tokens/sec.
All major labs are compressing frontier intelligence into cheaper/faster models rather than shipping new capability jumps. Sonnet is positioned as the workhorse for tool/computer use and autonomous agents, while Opus/Fable remain the "brain."
Brandon coined "defensive code scaling": as models get better at exploiting code, defensive review must scale proportionally once real money is involved.
Patrick predicts Anthropic will ship its own ambient agent within ~2 weeks (following Google Spark, Meta Muse, OpenAI Dots) — but warns against predicting competitive outcomes beyond a week out.
Ty argues OpenAI wins long-term via broad consumer reach, while Anthropic's developer focus serves a shrinking "coder" niche as builders shift toward orchestration. The group also debated whether AI labs should be public companies, citing Musk's 85% voting control at xAI as a healthier structure.
Agent workflows and infrastructure
Brandon Hancock built a personal SDLC skill that removes him from the coding loop: git work trees per task, interactive HTML planning converted into a requirements "contract," then a builder/reviewer loop (Opus builds, Sol reviews) producing an HTML proof document with GIFs.
Daniel Zivkovic orchestrates via Fable 5.1 in "SDLC Dark Factor" mode, delegating to cheaper models by task complexity, using a Codex plugin as a cheap peer reviewer, and converting overnight agent runs into Kindle ebooks for reporting.
Andrew Nanton's "Herder" tool lets different agent harnesses (Claude Code, Codex, etc.) communicate and be detached/reattached remotely.
Patrick rebuilt his personal infra into an "Agent Operated Environment": JSON-driven visual dashboard, three-node Proxmox mini-PC cluster, Mac Mini for local inference, Traefik + StepCA for internal HTTPS, Authentik for auth, Tailscale for access — and retired n8n entirely in favor of direct Python pipelines.
Minimum viable specs for running coding-agent CLIs in a VM: 4 vCPUs / 16GB RAM.
Cost and model selection
Brandon's cost comparison: a data-enrichment task cost ~14 hours/$400 via GPT-class models vs. ~$4 with DeepSeek V4 Flash — use cheap models to validate lead lists before building further. Instantly response rates scale sharply with audience specificity.
Tom Welsh's Cursor silently switched default models to Grok, causing a surprise $300 bill — watch your default model settings. He's likely moving to Claude's $200 plan.
Scott's Tight Code runs multiple AI engines (Opus/Sonnet security review, Codex adversarial review) on every GitHub push — not just PRs like CodeRabbit — with ~8% false positives over 2,000+ pushes, configurable at global/client/repo level.
Scott used Claude as a "statistician" to build performance-level curves from ~460,000 scraped race results. Brandon suggested turning the visualization into an email-capturing lead magnet and building the list before launch.
Scott's video pipeline chains ElevenLabs (voice/music/SFX), Higgsfield (images/animation), HeyGen (avatars), and Remotion for local orchestration.
Shipping and go-to-market
Ty Wells launched an ERP platform Sept 15 and onboarded ~1,500 warm leads using a T3 Code multi-agent orchestration system (reporting via Telegram) with no manual build work. He keeps a stable "core" system with per-customer customizations layered separately to avoid downtime risk.
Tom's revived "Asset MS" tool now sells to farmers at ~£5,000 per deployment for NFC-tagged equipment tracking.
Paul Miller: interactive prototypes convert buyers far better than slide decks. Back-office middle-management pain points are a growing market as companies can't afford that layer. NZ VCs reportedly ask for ~50% equity vs. ~8–12% in more evolved markets.
Brandon: engineers should flip the 80/90% product-time ratio and spend 90% of time with customers early on. New entrants should price low/free to build pipeline before raising prices. VC pitches are ~75% about a proven path to revenue.
Juan Torres: pricing power arrives when you're apologizing for raising prices due to demand. He's pivoting toward recurring corporate events (Christmas parties, proms) over one-off venue deals, targeting Halloween/Thanksgiving/Christmas as sequential stress-test launch windows. Cold outreach typically doesn't convert until the 2nd–3rd follow-up.
Juan's photo booth business: new OptiPlex-based serverless printer setup (Tailscale/WireGuard → EC2 → S3 → CloudFront), and country clubs beat wedding planners as customers since clubs directly control budget.
Patrick built a Claude-based internal "support skill" so employees get company-specific answers (e.g., how to open a ServiceNow ticket) instead of generic responses — and had an AI agent deliver his self-introduction at a new hire's onboarding.
Brandon on sales reality: government sales cycles are extremely slow (~5 months of bureaucracy vs. an ideal week-or-two), and he's currently drawing no income while pursuing an acquisition/fundraise that "could happen any minute."
❓ KEY Q&A
Q (Ryan C): Did I miss something — has OpenAI set up a new subscription tier?
A (Patrick Chouinard): Yes, Dev Day released a new $500 ultra-fast tier among many other things.
Q (Juan Torres): Is your plan in HTML instead of markdown to make it interactive?
A (Brandon Hancock): Yes — it lets him structure communication like a "boss" would want an initiative presented, with problem statement, ideal solution, UI mockups, and key decisions.
Q (Shakur Abdullah): How do you keep long agent runs going without babysitting rate limits during an 8-hour run?
A (Brandon Hancock): Spreads workload dynamically across Claude $200 / OpenAI $200 / Grok $100 plans based on capacity.
Q (Tom Welsh): Does the $200 Claude plan cover everything without running out of quota?
A (Brandon Hancock): It's thin margin; he runs all three subscriptions simultaneously and shuffles work; ran out once after heavy Astra 6 High usage.
Q (Tom Welsh): Are you all pushing Claude CLI instead of Cursor now?
A (Brandon Hancock): Yes, barely uses Cursor; mostly lives in CMUX-style multi-pane setups.
Q (Brandon Hancock): What exactly are you building for these client businesses (Ty's ERP)?
A (Ty Wells): Accounting/operations ERP customized per client (e.g., boat listings integrated with QuickBooks).
Q (Brandon Hancock): Slow nurture or straight conversion for bringing customers in?
A (Ty Wells): Warm leads — ~1,500 existing Global Tech Systems customers with shared pain points.
Q (Brandon Hancock): How do you handle downtime/versioning as customer data grows?
A (Ty Wells): Core system stays stable on its own DB; customer customizations are separate; downtime is communicated upfront (customers already tolerate this with tools like QuickBooks Online).
Q (Juan Torres): What process is being automated — upload a spreadsheet, get insights?
A (Ty Wells): Yes — system flags issues from data exports and self-onboards; hooks into common software directly or agents build a connector.
Q (Brandon Hancock): Was Higgsfield mainly used for voice in the High Rocks video?
A (Scott Rippey): No — ElevenLabs handles voice/music/SFX; Higgsfield does images/animation; HeyGen does avatars; Remotion assembles everything.
Q (Brandon Hancock): With OpenAI's subscription-sharing for inference, is that the confusing "double usage" messaging?
A (Patrick Chouinard): No — it's a separate feature letting developers avoid becoming inference resellers by using the end-user's own subscription.
Q (Ryan C): Isn't Scott's tool already powered by users' existing subscriptions?
A (Scott Rippey): Yes — no inference costs passed on; users bring their own plan or API key.
Q (Brandon Hancock): Is Sonnet worth using, given he's mostly used Fable/Opus?
A (Patrick Chouinard): Sonnet isn't a coding model — it's the workhorse for tool/computer use and autonomous agents, not Opus/Fable-class intelligence.
Q (Brandon Hancock): Did OpenAI demo the ultra-fast mode?
A (Patrick Chouinard): Yes, it's shown in the Dev Day video.
Q (Brandon Hancock): Are Dots using their own Decision API/classifier?
A (Patrick Chouinard): It appears to be a specialized, hyper-affordable/fast version of Luna rather than a separate external model.
Q (Brandon Hancock): How does OpenAI vs. Anthropic play out long-term?
A (Patrick Chouinard): Refuses to predict beyond about a week out given how fast plans/capabilities change.
Q (Brandon Hancock): Big Linux box with multiple VMs for your infra?
A (Patrick Chouinard): No — three Proxmox nodes on mini PCs; scales by adding more mini PCs.
Q (Juan Torres): What's your CloudFront equivalent for the lab?
A (Patrick Chouinard): Tailscale connecting to Traefik as reverse proxy, which also hosts StepCA for internal HTTPS.
Q (Brandon Hancock): What is Traefik?
A (Patrick Chouinard): The reverse proxy — routes like forge.patchoutech.lab map to internal IP/port.
Q (Brandon Hancock): What's the conversion rate on cold outreach campaigns?
A (Brandon Hancock): Depends on offer/targeting specificity; could be 5–15%+; "conversion" here means a response, not a sale.
Q (Brandon Hancock): Why country clubs specifically?
A (Juan Torres): They directly control budget and pay for events themselves, unlike wedding planners who are referral sources; also higher-end clientele.
Q (Brandon Hancock): What prices are you quoting for these events?
A (Juan Torres): $1,295 plus travel fee; the director liked it but hasn't responded in about a week.
Q (Juan Torres): What's a good VC proof of concept?
A (Brandon Hancock): Showing a direct/proven path to revenue is ~75% of it; the rest is scalability plus matching investor capital with founder time/execution.
Q (Brandon Hancock): Is the new investor situation with Paul happening?
A (Paul Miller): Yes — he wants to keep generating ideas and partner with capital for fast 3–4 month build-and-sell cycles.
🛠️ TOOLS AND CONCEPTS MENTIONED
OpenAI Dev Day releases — GPT-6.1 Sol, the "Dots" ambient agent, GPT Space, a Decision API built on Luna, and a $500/mo Cerebras ultra-fast tier (~3,000 tokens/sec).
Anthropic Opus 5.5 / Sonnet 5.5 / Fable 5.1 — Anthropic's recent model lineup, discussed head-to-head against OpenAI's releases.
Claude Code — Brandon's daily-driver coding harness, including the Advisor slash command.
Codex / Codex CLI — used for builds, with "Astra Light" acting as a peer reviewer for adversarial and security review.
Sol 6 / Sol 6.1 — builder and reviewer models inside Brandon's SDLC loop.
DeepSeek V4 Flash — a low-cost LLM used for lead classification.
Lavish / High Throughput Agentic Engineering — an interactive planning protocol shared by an ex-Meta L8 engineer.
Deep Plan / Deep Proof — Brandon Hancock's custom skills for planning and proof reporting.
IndieDevDan (YouTube) — a source of inspiration for agentic workflows.
Herder / Herdr — Andrew Nanton's tool for cross-harness agent communication.
T3 Code — Theo's mobile-accessible, multi-project, cross-harness coding harness; an OpenAI launch partner.
CMUX — a multi-pane terminal environment for running agents in parallel.
CC Blackbox (ccblackbox.app) — Scott Rippey's free Mac/iOS Claude Code session manager with iPhone remote control, SQLite storage, iOS Simulator, and Vercel account connections.
Tight Code (tightcode.dev, promo code BRANDON20) — Scott's multi-engine automated code-review tool that triggers on every GitHub push, pairing Opus/Sonnet with Codex for adversarial and security review.
CodeRabbit — a PR-based review tool mentioned as a comparison point.
video-generator repo (github.com/scott-rippey/video-generator) — Scott's Claude Code/Codex-compatible video generation tool.
ElevenLabs / Higgsfield / HeyGen / Remotion — Scott's video pipeline: voice/music/SFX, images/animation, avatars, and local assembly.
Fable 5.1 — Daniel Zivkovic's orchestrator model.
Instantly (instantly.ai) — cold email outreach tool.
Dokploy — the deployment platform Paul Miller uses to auto-deploy POCs.
Proxmox — virtualization platform powering Patrick's and Paul's home labs.
Ollama / LM Studio — local inference running on a Mac Mini.
Traefik / StepCA / Authentik / Tailscale — Patrick's stack for reverse proxy, internal certificates, authentication, and secure network access.
Prometheus / Grafana — observability and monitoring stack.
NATS / Slipstream — internal messaging system.
DSPy — a framework for automatic prompt improvement, covered in Daniel's meetup.
Limitless — Ty Wells' AI recording device for client meetings.
Asset MS — Tom Welsh's NFC-based asset tracking tool.
QuickBooks Online — accounting integration and reference point in the discussion.
Kindle — used to convert agent-run summaries into ebooks.
Telegram — the status reporting channel for Ty Wells' agents.
AWS EC2 / S3 / CloudFront — backend for Juan Torres's photo booth print pipeline.
OptiPlex PC — Juan's repurposed serverless printer server.
Canon Selphy CP1500 — compact wireless photo printer suggested by paulmiller.
God's Eye view (github.com/bilawalsidhu/gods-eye-view) — a "consumer Palantir"-style world-monitoring tool, discussed in relation to Patrick's dashboard concept.
Home-lab build parts (Paul Miller) — ASUS ProArt B850-Creator WiFi Neo motherboard, AMD Ryzen 9 9950X, Cooler Master MasterLiquid Core Nex 360 ARGB cooler, be quiet! Dark Power 14 1200W PSU, and a Phanteks Enthoo Pro II Server Edition case.
📎 SHARED RESOURCES
Here are all the links and resources shared during the call:
Videos and Interviews
https://www.youtube.com/watch?v=MSbacZ99E14 — "High Throughput Agentic Engineering with ex-Meta L8 Principal" (Brandon's inspiration video)
https://www.youtube.com/watch?v=_VyF0naZ4AI — Mustafa Suleyman (Microsoft AI CEO) interview (Tom Welsh)
https://www.youtube.com/watch?v=mSjaMyP5QjY — Modern Wisdom podcast panel on AI/society (Scott Rippey)
https://youtu.be/c9WCcD4fH6c — video shared by Brandon Hancock (JEV/classification discussion)
https://www.youtube.com/watch?v=4RmPNxhfqHg — Serverless Toronto meetup recording with slide deck (Daniel Zivkovic, DSPy-related)
https://www.youtube.com/@peterdiamandis — Peter Diamandis YouTube channel (Brandon Hancock)
Articles and Tools
https://claude.com/blog/the-ai-native-sdlc-playbook — Daniel Zivkovic's shared article on the disappearing "dev phase" of SDLC
https://ccblackbox.app/ — Scott Rippey's CC Blackbox app
https://tightcode.dev/ — Scott Rippey's code-review product (code BRANDON20)
https://github.com/scott-rippey/video-generator — Scott's video generator repo
https://github.com/bilawalsidhu/gods-eye-view — "God's Eye" world-monitoring project (Biggi Fraley)
https://instantly.ai/ — cold email tool (paulmiller)
https://www.linkedin.com/in/magmainc/ — Daniel Zivkovic's LinkedIn
Paul Miller's PC Build
https://www.asus.com/us/motherboards-components/motherboards/proart/proart-b850-creator-wifi-neo/ — motherboard
https://www.coolermaster.com/en-global/products/masterliquid-core-nex-digital-series.html — CPU cooler
Miscellaneous
https://au.pcmag.com/printers/97089/canon-selphy-cp1500-wireless-compact-photo-printer — printer suggestion for Juan's photo booth (paulmiller)
🔄 FOLLOW-UPS WORTH EXPLORING
Open questions and follow-ups to revisit next week:
Brandon Hancock to share his "Deep Plan" and "Deep Proof" skill files with the group.
Tom Welsh to likely upgrade from Cursor's basic plan to Claude's $200 plan after an unexpected $300 charge.
Brandon Hancock to try Sol 6.1 as a potential orchestrator model.
Ty Wells to travel to install four more ERP systems, reaching 10 installs by Sunday.
Brandon Hancock to experiment with staged/versioned database migration approaches and share findings with Ty Wells.
Scott Rippey to provide promo code "Brandon20" for discounted access to his code-review product and drop links (video pipeline, CC Blackbox, code-review tool) into chat.
Brandon Hancock to try Scott's tools on a future non-regulated side project.
Scott Rippey to pursue outreach for his security-review tool, including engaging on Reddit.
Patrick Chouinard to send Brandon a screenshot of his AOE visual dashboard once built.
Patrick Chouinard's team proceeding with a Claude rollout for a 2,000-person company, with contingency plans for potential breaking platform changes.
Brandon Hancock to attempt launching his YouTube lead-magnet product using the validated lead list.
Juan Torres to implement a structured 3–5 touch outreach cadence for country clubs/venues.
Juan Torres to field-test the new printer feature at upcoming events.
Juan Torres to test rev-share, broker/agency, and direct-to-business models in parallel via a few deployed units before scaling.
Brandon Hancock to get the name of an entertainment-broker company model for Juan's business.
Juan Torres to lower/match pricing to classical photo booth rates with "founder discount" framing.
Juan Torres to pivot outreach toward recurring corporate events (Christmas parties, proms) over one-off venue deals.
Juan Torres to target Halloween as a 30-day stress-test launch window, followed by Thanksgiving and Christmas pushes.
Juan Torres to deploy the app free at social events via San Diego tech contacts for word-of-mouth leads.
Juan Torres to deprioritize the portable hardware build this month in favor of sales.
Paul Miller to send Juan business strategy ideas and offer a VC introduction.
Patrick Chouinard and Paul Miller to schedule a call to swap insights on infra vs. "build-sell-repeat" venture models.
Brandon Hancock to continue working unpaid until an acquisition/fundraise closes, then resume more regular call participation.
1
3 comments
RecapFlow : September 29nd 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
+2
2
+2
3
+1
4
+1
5
+1
Powered by