16d • General discussion
RecapFlow : October 6th Coaching call analysis
📝 SUMMARY
This call captured the community at its best: an informal, rotating mix of peer coaching, technical show-and-tell, and open Q&A where members traded real-world lessons on AI tooling, infrastructure, and side projects. Two big themes ran through the session. First, moving beyond slow sequential LLM loops toward typed, parallel-evaluation models (GEV/JEV via TypeSafe) for classification and judgment tasks, with concrete use cases from council vote tracking to forensic case analysis. Second, building and hardening real products: a security-rebuilt digital signage SaaS, a Supabase-to-Convex migration, an open-source CRM, a HIPAA-sensitive forensic-psychiatry tool, and Patrick's impressive setup deploying Claude-based support skills to ~2,000 enterprise users alongside a self-diagnosing personal agent. Whether you came for the architecture debates, the one-prompt side builds, or the school-event AI booth preview, there was something practical to take away — and plenty of the friendly banter that makes these calls worth showing up for.
💡 KEY INSIGHTS
Patrick Chouinard used ChatGPT for an election decision by describing leadership priorities only (not parties) and asking for the government configuration that maximizes them — found it "surprisingly accurate."
Paul Miller: dual-GPU home servers splitting model loads are cheaper than one large GPU.
Paul Miller: for GEV/JEV-type models, best practice is breaking large questions into many small ones routed to appropriately-sized LLMs rather than one big complex prompt.
Paul Miller: JEV is useful for categorizing large corpora (e.g., council documents) cheaply to detect votes/topics.
Paul Miller: JEV beats a standard loop because it evaluates a batch of typed questions in parallel against the same state in one call — no sequential latency and no "context rot" from one question bleeding into another; adding questions barely changes response time.
Paul Miller's own approach: "TypeSafe first, then train my own model."
Patrick Chouinard is interested in extending GEV/JEV-style evaluation toward agentic "state change" monitoring over time.
Ryan C rebuilt his signage SaaS with Claude + a security-review tool after the original Sonnet-built code had ~300 security/quality findings.
Ryan C built an 8-step onboarding visualizer using Firecrawl to pull brand assets for auto-generated mockups.
Ryan C: commercial Android signage players are now price-competitive with Raspberry Pis and better for continuous video playback.
Ryan C recommends the "advisor" command to pair Fable 5.1 as an advisor to Opus 5.5 ("5.5 kicks the shit out of 5").
Morgan migrated a client app from Supabase to Convex specifically for automatic sync.
Morgan sees little quality difference between Fable and Opus beyond tone, suspecting his own legacy "skills" may now hold back newer models.
Patrick Chouinard: using fewer custom skills/prompts is yielding better results as base models improve.
Shakur Abdullah is building an open-source CRM ("Remold," on Convex) designed to reuse users' own subscriptions instead of incurring API costs.
Shakur Abdullah tested the AI agent "Dot," which proactively flagged pending invoice payments unprompted.
Shakur Abdullah is drafting a book ("Hand It Off") as an "Automate the Boring Stuff"-style guide for AI-driven automation.
Andrew Nanton: JEV is cheap/fast enough that he's actively seeking more use cases; behaves unlike standard LLM prompting.
Andrew Nanton: TypeSafe/JEV offers Zero Data Retention via direct sales request, not obviously documented.
Andrew Nanton built a forensic-psychiatry tool with citation linking, Bates numbering, transcript-to-audio playback, and Typst export.
Andrew Nanton is experimenting with A2UI to let agents dynamically generate task-specific UI components rather than static interfaces.
Andrew Nanton uses GCP/Azure with signed BAAs for HIPAA compliance since local models aren't yet strong enough for complex analysis.
Andrew Nanton: a £10k Mac Studio with 500GB RAM could run local models for his use case, but he's holding out for Apple's M7 (rumored ~2027) as a major step forward for local LLM performance, though "ruinously expensive."
Patrick Chouinard uses GitHub as a "production platform" — enterprise Claude plugins sync directly from repos into the Claude marketplace via PR-based auto-merge, a mental shift from standard review workflows.
Patrick Chouinard: giving agents an infrastructure "inventory" lets them autonomously diagnose/fix problems — his DOTS agent found and cleaned 74GB of misplaced Ollama files and stale NPM cache unprompted.
Patrick Chouinard: a side dashboard project revealed that hosted honcho.dev could be self-hosted on his Proxmox cluster, and the agent implemented the fix itself.
Juan Torres: agentic monitoring of staging (not just production) environments is increasingly necessary, since cache/storage overflow can silently propagate to production.
Patrick Chouinard: customizing community tools (e.g., Matt Pocock's "Grill with Docs") with enterprise-specific pre-answered context cut question counts from ~100 down to 10–20.
Morgan: a developer (Jason) built a UI that syncs with Claude via a shared JSON state file instead of API calls, running entirely within a Claude subscription without API costs.
Morgan reiterated Matt Pocock's philosophy of keeping single-purpose AI-built tools "as simple and small and stupid as possible."
Patrick Chouinard: Claude's new "Mods" feature let him build and deploy a thumbs-up/down feedback mechanism (anonymized logging) to 10 users within half a day.
❓ KEY Q&A
Q (Paul Miller): Is the provincial voting system fast like a 24-hour result, unlike NZ/Germany/Israel?
A (Patrick Chouinard): Full count takes longer, but the outcome was known ~45 minutes after polls closed; it's first-past-the-post, not proportional.
Q (Patrick Chouinard): Do you still see a valuable difference using Fable vs. Opus?
A (Morgan): No real difference besides tone; his cleanup "skills" may be masking true model differences.
Q (Shakur Abdullah, chat): So how is it better than a standard software program?
A (Paul Miller, chat): JEV takes state plus a batch of typed questions in one call, evaluating each in parallel and isolation — faster than looping calls and avoids one question muddling another ("context rot").
Q (Shakur Abdullah, chat): What model did you use to build it?
A (Paul Miller, chat): "TypeSafe first then train my own model."
Q (Patrick Chouinard, re: Shakur's CRM): How would you like us to contribute?
A (Shakur Abdullah): It's fully open source; the repo was private, so interested people should ask to be added, fork it, and make changes.
Q (Shakur Abdullah): What is this Google Cloud thing — can it run on a small old box, and how is confidentiality handled?
A (Andrew Nanton): Hosted on GCP (also tried Azure); has BAAs with both and uses their hosted models (Gemini, Google-hosted Claude) since local models aren't yet strong enough for heavier analysis, though simple classification could move local soon.
Q (Shakur Abdullah, chat): When are they (Apple) saying the M7 will arrive?
A (Andrew Nanton, chat): 2027 — "undoubtedly ruinously expensive."
Q (Morgan, re: coworker's question): Is there a way in Claude Desktop to give thumbs-up/down feedback on a skill or response?
A (Patrick Chouinard): No native feature, but Claude's new "Mods" hooks layer let him build a custom feedback mod logging anonymized prompt/skill data — deployed to 10 users within half a day.
Q (Morgan): What codebase/language is ScratchMD written in?
A (Patrick Chouinard): He didn't ask — just told the AI to build and deploy it for Mac and let it decide.
🛠️ TOOLS AND CONCEPTS MENTIONED
TypeSafe / GEV/JEV models – typed, parallel-evaluation model architecture used by Paul, Patrick, and Andrew for cheap, fast categorization/classification and ZDR-compliant inference.
T3 server code – reference implementation for self-hosted LLM infra without extra API costs.
Opus 5.5 / Opus 5 – model versions compared by Paul, Ryan, Morgan, and Patrick.
Fable (incl. Fable 5 / 5.1) – AI coding assistant used by Morgan and Ryan; paired as "advisor" to Opus 5.5.
Convex – backend used by Morgan (client migration, time-tracking app) and Shakur (Remold CRM) for auto-sync.
Supabase – prior backend Morgan migrated away from.
Claude (Claude Code, Claude Desktop, Claude Chat/Cowork) – used across the call for app builds, enterprise skills, and plugin publishing.
Claude Mods – new hooks/interaction layer used by Patrick to build a feedback mechanism.
Tight code (Scott's tool) – security auditing tool Ryan used to fix Sonnet-generated code.
Sonnet 4.2 – original model used to build Ryan's signage app, later found insecure.
FFmpeg – used by Ryan for reviewing screen-recorded design inspiration.
Firecrawl – used in Ryan's visualizer to pull brand assets from client websites.
Raspberry Pi (4 and 5) – signage driver hardware (Morgan's "SignPi"), compared to Android players.
Commercial Android signage players – adopted by Ryan over Pi/Windows for cost and reliability.
Dot – AI agent Shakur tested that proactively surfaced invoice tasks.
GrokBot, Instinct – other agent platforms Shakur is considering.
Typst – typesetting system used by Andrew for legal document export.
A2UI – standard Andrew uses for agent-generated dynamic UI components.
CodeMirror – editor framework underlying Andrew's document interface.
GCP / Azure – cloud hosting with BAAs for HIPAA compliance (Andrew).
Turso – distributed SQLite service Andrew is evaluating for local-first storage.
Gemini / Google-hosted Claude models – used by Andrew for privacy-compliant inference.
GitHub / GitHub Copilot – used as production sync target for Claude plugin skills.
Azure DevOps – internal repo preceding GitHub build pipeline (Patrick).
ScratchMD – Patrick's custom native Mac markdown scratchpad with Mermaid support.
Operator (VM) / NATS / Invisible / DOTS / Codex – components of Patrick's personal agent orchestration framework.
Ollama – local model runner; DOTS fixed a misconfigured storage location.
Honcho.dev – agentic memory management service; Patrick self-hosted it on Proxmox.
OpenRouter – routing layer in Patrick's self-hosted setup.
Hermes – Patrick's Discord-connected infra notification system.
Agent Ops – part of Patrick's monitoring stack (noted as "not stable").
Proxmox – Patrick's home-lab virtualization cluster.
GPT-5/6 "Luna" – used by Patrick to convert meeting transcripts into build prompts.
Matt Pocock's "Grill with Me" / "Grill with Docs" – community skill framework Patrick customized for enterprise use.
"Grill With UI" (Jason) – skill syncing a UI with Claude via a JSON state file instead of API calls.
Clockify – commercial time tracker Morgan replaced with a custom Convex app.
Drafts – Mac markdown/text capture app recommended by Andrew.
tot.rocks – another lightweight capture app Andrew recommended.
Netgear hotspot / Duo Mini antenna – networking hardware for Juan's AI Booth Studio deployment.
Multica.ai – team devops workflow tool Paul referenced to Patrick.
Screenli.co.uk – example signage SaaS site referenced by Patrick.
📎 SHARED RESOURCES
Links & Resources Shared
Products & tools mentioned in chat banter:
https://www.neutonic.com/en-us/ — energy drink (Brandon Hancock)
https://ketone.com/ — energy drink (Scott Rippey)
AI meeting notes:
https://fathom.video/customize — Fathom notetaker settings
https://zoom.us/oauth/authorize?client_id=ULKDU8_8TLySSz8ZMyO9WQ... — Otter.ai Zoom connection invite
Member demos & projects:
https://screenli.co.uk/ — example signage SaaS (Patrick Chouinard)
https://multica.ai/ — team devops workflow tool (Paul Miller)
https://github.com/shakur-cmd/remold — Remold CRM open-source repo (Shakur Abdullah)
https://honcho.dev/ — agentic memory management service (Juan Torres)
https://github.com/jasonku09/grill-with-ui — "Grill With UI" skill (Morgan)
Apps for capture & notes:
https://getdrafts.com/ — Drafts Mac app (Andrew Nanton)
https://tot.rocks/ — lightweight capture app (Andrew Nanton)
🔄 FOLLOW-UPS WORTH EXPLORING
Paul Miller will share findings from his JEV-based council-document categorization project with the group.
Paul Miller will post YouTube examples of GEV/JEV model usage and best practices in the forum.
Shakur Abdullah will add anyone requesting access to the Remold CRM repo so they can fork/contribute (fulfilled in-call by posting the GitHub link).
Morgan will share photos of his "Builds with Morgan" branded merch once received.
Andrew Nanton will look into Convex and Turso as potential local-data solutions for his confidentiality-sensitive application.
Patrick Chouinard will finalize and deploy Claude-based support skills to ~2,000 enterprise users before go-live.
Patrick Chouinard is publishing a GitHub-based plugin distribution system across Claude Chat/Cowork, GitHub Copilot, and Claude Code.
Patrick Chouinard will continue rebuilding his personal agent framework (Operator, NATS, DOTS) and share results with the group.
Juan Torres will deploy the AI Booth Studio at a school event, including network/hotspot and printing tests.
Juan Torres will attempt a custom "Cougar" mascot style with reactive expressions and record reactions at the event.
Juan Torres will share screen-recorded outputs from the event as early as next week, confidentiality permitting.
Morgan posted Jason's "Grill With UI" skill link to the group in-chat.
Andrew Nanton re-posted the Drafts app link in chat after it got lost earlier.
0
0 comments
RecapFlow : October 6th 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
+3
2
+3
3
+2
4
+2
5
+1
Powered by