RecapFlow : October 13th Coaching call analysis

📝 SUMMARY

This week's call was a wide-ranging mix of infrastructure talk, business updates, and honest discussion about what it actually takes to scale AI usage. Members swapped notes on managing multiple AI subscriptions and their costs, Patrick Chouinard walked through leading a large enterprise Claude rollout at a Canadian financial institution (including the token-economy culture shift and employee confusion between models and "harnesses"), and Ty Wells shared progress on his Island Flow ERP/accounting product for Bahamas businesses plus a demo of a redesigned T3 interface with draggable widgets. Juan Torres reported on deploying his AI photo booth at a high school homecoming — Wi-Fi interference near a military base included — and the feedback that he needs a proper landing page to attract institutional clients. The group also dug into technical territory: Patrick's NTFY-based secure push-notification setup tied to Tailscale, his "AOE" (Agent Operated Environment) project, Ty's multi-machine Proxmox setup, and ChatGPT's "Dots" feature. A recurring theme throughout: the gap between AI power users running multi-agent setups and "normies" just getting started, and how to navigate the cost, confusion, and adoption challenges in between.

💡 KEY INSIGHTS

Ty Wells (business owner) frames his ~$2,000/month in AI subscriptions as cheaper than hiring "a senior software architect, engineer, QA tester," justifying heavy usage.

Brandon Hancock noted that after Codex's usage limits were cut in half, the subscription "isn't that good anymore," and he has been timing resets to maximize usage.

Patrick Chouinard described enterprise Claude rollout dynamics: personal AI subscriptions are blocked entirely at his financial-institution client due to data privacy/audit requirements — all usage must go through the company tenant.

Patrick observed a cultural shift at that client: employees used to negotiate over return-to-office days; now the key negotiating point is "which model can we access," with people requesting more tokens instead of overtime or pay.

Patrick flagged a major enterprise education challenge: users often don't distinguish between a model (e.g., Claude Opus) and the "harness" running it (Copilot vs. the Claude app vs. a Word plugin), leading to inconsistent experiences and misplaced judgments about model quality.

Paul Miller remarked (unconfirmed, anecdotal) that different AI companies' models show different self-ranking behavior — Claude reportedly ranks itself lower against competitor models, while OpenAI's models behave differently.

Shakur Abdullah's "game of horse" experiment pitted AI models against each other in designing and scoring competitive tasks (coding, research, video, etc.): a 5.5 model reportedly won overall, beating Fable and Astra, with Soul 6.1 becoming the best helper for large tasks; smaller models (Opus, Haiku) underperformed. One model (Fable) accused another (Soul) of "cheating" by inventing scoring criteria mid-game — an emergent, unplanned dynamic Shakur found noteworthy.

Ty Wells described a tiered business philosophy for Island Flow: full ERP infrastructure (payroll, time/attendance, invoicing, etc.) sits underneath lightweight "skins," so a single-operator seller can start minimal and grow into enterprise-level features without migrating platforms.

Ty distinguished his onboarding approach: new businesses get a standardized chart of accounts, but customers migrating existing data use a dedicated "onboarding agent" that checks data and flags issues — and only the customer decides what to keep, discard, or clean up, framed as "a good time to clean it up."

Patrick noted that in his enterprise deployment, mass basic-skills training is now handled by an external training company due to scale, so his focus has shifted to specialized/"fringe" skill-building, aided by an automated training generator that is "working pretty well."

Patrick mentioned talk in the T3 code GitHub repo about integrating a "mod" concept natively into T3 — Ty, a T3 user who forked the codebase, is already informally building toward that.

Patrick built a personal FinOps dashboard aggregating usage/spend across Claude, Codex, Open Router, Eleven Labs, Firecrawl, Honcho, and Tavali (with plans to add GitHub), using Prometheus/Grafana for monitoring and a Traefik-style reverse proxy for front-end routing; Opus reportedly took ~14 hours to build it.

Patrick shared that Claude, while helping build his FinOps/Agent Ops infrastructure, proactively flagged that he lacked a push-notification service and suggested installing NTFY — an unprompted infrastructure recommendation.

Patrick explained NTFY was chosen specifically because it can stay within his home network via Tailscale, avoiding exposing a sensitive re-registration link through Discord, email, or any service that leaves his environment.

Patrick now separates local development from deployment: he does a full build deployed to a production server rather than running raw code from the dev machine, so missing credentials on the dev machine aren't a problem.

Patrick uses a dedicated secret-management server (referred to in the transcript as "physical"/"Infosil," likely Infisical — name uncertain) so credentials are never stored in local keychains across his machines.

Patrick's AOE (Agent Operated Environment) project is described as not a documentation project but an "AI context project" — mostly Markdown/YAML/JSON — acting as a master inventory of all machines, applications, and which sub-project manages each, letting T3 code navigate and modify any project from a single thread.

The AOE system automatically updates network diagrams in Excalibraw whenever inventory changes, and tracks where each project is cloned or deployed so updates propagate to the correct machines.

Cross-project communication happens through GitHub Issues; ChatGPT's "Dots" feature can read across these projects, flag duplicate issues, and offer to create higher-level issues or reorganizations.

Patrick said he loves Dots because it required no configuration and proactively found things to fix, though he still values Hermes for things Dots cannot do.

Patrick and Ty both keep separate paid AI subscriptions for personal vs. work use: Patrick uses one 20X Claude and one 5X Codex subscription for personal work, while Ty bills some of his five Anthropic and five Codex subscriptions to his companies.

Ty uses scheduled jobs in T3 code to automatically handle recurring credential-refresh actions (e.g., WhatsApp/Meta tokens) instead of relying solely on calendar reminders.

Patrick emphasized that getting multiple T3 code sessions/threads to talk to each other — so a specialized "home server" project with SSH keys can handle infrastructure while a frontend project stays unaware of secrets — was what drew him to T3 code.

Paul observed that his CTO (previously against AI) has become highly engaged, running multiple AI agents 24/7 to build pipeline features — to the point that Paul now needs a day to process what was presented.

Paul reflected on the large gap between early AI adopters and "normies," citing a trade-show sign supplier in New Zealand whose business lacked basic e-commerce/AI-enabled ordering capability.

Patrick advised Juan Torres that attracting corporate/institutional clients (schools, colleges) rather than just social-media-driven individual customers requires a proper landing page/reservation website, since procurement departments won't sign contracts via Instagram.

Juan Torres reported that his AI photo booth product's physical printer performed reliably under production conditions at a large event, despite Wi-Fi/radar interference challenges near a military base.

❓ KEY Q&A

Q (Shakur Abdullah): Is no one at Patrick's company bringing their own personal AI subscription to work?

A (Patrick Chouinard): No — personal subscriptions are blocked entirely because data would go into personal accounts instead of staying within the company's tenant; for data privacy and audit reasons (especially as a monitored financial institution), only business subscriptions can be used.

Q (Shakur Abdullah): Is that restriction specifically because Patrick's company is in finance?

A (Patrick Chouinard): Yes — large financial institutions that are monitored and audited cannot allow that kind of personal-account usage.

Q (Shakur Abdullah): Are employees at Patrick's company allowed to work remotely?

A (Patrick Chouinard): Yes, remote work is allowed, though currently the policy is RTO (return to office) four days a week, flexible; full-time remote work is not preferred by the company.

Q (Brandon Hancock): Was the Claude rollout announcement more like a big "launch/pitch" demo, or just a simple notification that the tool was available?

A (Patrick Chouinard): It was essentially just a notice (e.g., in the elevator and on an intranet banner) saying people could now request a Claude license via a website — not a big demo event — yet it still drove more than half the company to request licenses within the first half-day.

Q (Brandon Hancock): Is the general licensing strategy that everyone gets a budget but can request more if doing good work?

A (Patrick Chouinard): Not exactly tied to "good work" — there's a justification process required to request additional budget, and a ceiling exists; exceeding the ceiling requires escalation to higher-level decision-makers, though they hope not to hit that scenario given how the pilot was scoped.

Q (Paul Miller): Were employees mixing up or confusing which underlying model they were using within Claude (e.g., Haiku vs. Opus vs. Fable)?

A (Patrick Chouinard): Yes — because the same model (e.g., Claude) appears across multiple surfaces (Copilot, the Claude app, Word extension), users often don't realize they're interacting with the same model through different "harnesses," leading to confusion; some dismissed Claude as unsatisfactory after trying it only within Copilot rather than the native Claude app.

Q (Shakur Abdullah): How is the uptake going on Patrick's automated training modules that he spent time fine-tuning?

A (Patrick Chouinard): He's now focusing more on specialized/fringe skill-building training since a separate company now handles basic training (too many people for just two trainers); the automated training generator itself is working pretty well.

Q (Shakur Abdullah): Does Ty's Island Flow system integrate with Shopify?

A (Ty Wells): No — Island Flow is designed to replace Shopify, providing its own simple web/e-commerce "skin" for small sellers while the full ERP backend (inventory, customers, etc.) runs behind it.

Q (Shakur Abdullah): Can Patrick share the push-notification tool/product he mentioned Claude recommended?

A (Patrick Chouinard): Yes — it's called NTFY, an open-source product that can be installed as a VM or container locally; it communicates with a mobile app to push notifications, and in his setup it runs entirely through his own VPN/network so tokens never leave his network.

Q (Ty Wells, to Patrick Chouinard): Are you sim-linking your projects, or do you have a document, or does it just have access to everything?

A (Patrick Chouinard): It has access — the main AOE project contains the full inventory of machines, applications, and which sub-project manages them, but doesn't itself know how to interact with machines (a separate home server project handles that).

Q (Ty Wells, to Patrick Chouinard): Are you publishing to that environment, or is it polling that environment for inventory updates?

A (Patrick Chouinard): Both — it's read/write into the environment; every inventory change also triggers an automatic redraw of network diagrams in Excalibraw.

Q (Ty Wells, to Patrick Chouinard): Are you finding that's the [Dots] using it [across multiple GPT accounts]?

A (Patrick Chouinard): I only have a single ChatGPT account, so Dots operates from that one account's memory and history.

Q (Paul Miller, to Juan Torres): Have you had any thoughts on how you can make things a lot more portable with what you do and what you could do?

A (Juan Torres): He hadn't yet looked at a diagram Paul had sent, calling portability a hardware engineering task he still needs to tackle; he explained he currently favors a more elaborate, attention-grabbing setup (big screen, craftsmanship) over simpler alternatives like two standalone iPads because it draws more attention and vendor interest at large events.

Q (Patrick Chouinard, to Juan Torres): Have you started thinking about having a web front end for your business/product/services?

A (Juan Torres): Yes, he has an Instagram presence which he considers more important for his current clientele, but acknowledges he needs a landing page/reservation system, especially to attract higher-paying institutional clients; he has been delaying it due to being overloaded with event work.

Q (Patrick Chouinard, to Paul Miller): Have you discovered how to have multiple T3 threads talk to each other?

A (Paul Miller): Yes — he has a high-level "chief" session talking to multiple T3 sessions, aided by a fairly capable API that lets sessions instruct each other on how to communicate.

Q (Patrick Chouinard, to Paul Miller): What have you done this week?

A (Paul Miller): His CTO has been building multiple AI agents working around the clock on the product pipeline, which is pushing Paul to shift from driving development to being more proactively sales-focused; he also described a hardware/network upgrade effort (fiber to 2 gigabit, core switch, server rack) and an unrelated trade-show sign procurement challenge in New Zealand highlighting how far behind "normies" are in e-commerce/AI adoption.

Q (Shakur Abdullah, in chat): Have you tried Grok bot?

A (Paul Miller, in chat): Yes but trust with Grok a challenge.

Q (Shakur Abdullah, in chat): Do you have multiple claude? ... Ty what are you doing with all of them?

A (Ty Wells, in chat): Building my ERP platform and my other business systems and some side projects like my golf app but mostly each customer has their own agent for onboarding so that's where it's spread out.

🛠️ TOOLS AND CONCEPTS MENTIONED

Codex — AI coding subscription; usage caps were cut in half, reducing value per Brandon Hancock.

Claude (subscription/app) — Primary enterprise AI tool rolled out at Patrick's financial-institution client; also used for personal projects (portfolio site redesign, FinOps dashboard build); Patrick uses a 20X subscription.

GLM, DeepSeek — AI subscriptions Ty Wells mentioned having among his "five and five" stack.

CLI Proxy — Tool mentioned by Patrick Chouinard (via Theo's T3Code videos) for managing/switching between multiple AI subscriptions; Theo advised not using the default setup.

T3Code (T3) / T3 code — Coding/orchestration tool/platform used by multiple members (Patrick, Ty, Paul) for distributed development work and multi-session/project communication; has planned "mod" integration discussed in its GitHub repo.

Claude mods — Customization layer for Claude desktop client; Patrick built a collapsible status-line mod and a thumbs-up/thumbs-down feedback-capture mod.

Island Flow — Ty Wells' ERP/accounting SaaS product launched for Bahamas businesses, with onboarding agents and customizable "skins."

Shopify — Referenced as a comparison point; Island Flow's e-commerce "skin" is positioned as a Shopify replacement.

QuickBooks — Mentioned as an easier data source for onboarding into Island Flow compared to "obscure" systems or spreadsheets.

Apple Watch / Garmin — Ty Wells built a personal golf-tracking app on Apple Watch after his Garmin device had charging/connection issues.

Copilot — Enterprise AI tool already deployed company-wide at Patrick's client prior to the Claude rollout, causing model-confusion among users.

Prometheus / Grafana — Monitoring stack used in Patrick's personal FinOps dashboard.

Open Router, Eleven Labs / 11Lab, Firecrawl, Honcho, Tavali — API/service providers aggregated into Patrick's personal FinOps usage dashboard.

NTFY — Open-source local push-notification service Claude suggested Patrick was missing; runs via VPN/Tailscale so data stays on his network. (https://ntfy.sh/)

Opus / Opus 5.5 — Model used to build Patrick's FinOps dashboard (~14 hours); Opus 5.5 also referenced by Paul Miller as part of his current toolset.

Tailscale — VPN used by Patrick and Ty to securely connect to home servers/MacBook without exposing secrets externally.

Proxmox — Virtualization platform on Ty Wells' hardware used to distribute T3 code jobs across machines.

Infisical / "physical"/"Infosil" (name uncertain) — Secret management service used by Patrick (and attempted by Ty) to centralize credentials instead of local keychains.

Telegram — Used by Ty Wells for receiving push notifications/links, though it gets cluttered with many messages.

Discord / WhatsApp / email — Messaging channels Patrick explicitly avoided for the secure NTFY use case; Ty uses WhatsApp tokens that require periodic refresh via Meta.

GitHub — Hosts Patrick's shared organization (including Hermes's separate account) and is used for cross-project issue-based communication.

Excalibraw — Diagramming tool automatically updated by Patrick's AOE system to redraw network/infrastructure diagrams.

ChatGPT "Dots" — ChatGPT feature/agent Patrick uses that has access to GitHub, memory, and can proactively flag and offer to fix issues across projects.

Muse ("wonderful model") — Newly mentioned AI model Patrick asked about; name/details uncertain, Ty had only briefly seen mention of it.

Ubiquiti — Networking hardware brand Paul Miller ordered from while upgrading his home network/core switch.

OpenAI real-time (voice) — Used by Ty Wells for a voice interface ("voice two") with ChatGPT-like access to his agent operations.

ACP — Protocol/integration Patrick mentioned as providing basic integration between T3 and Hermes, which he hopes to complete to fully drive Hermes.

Grok bot — Mentioned in chat by Shakur Abdullah; Paul Miller responded that trust with Grok is a challenge.

AOE (Agent Operated Environment) — Patrick Chouinard's custom AI-context repository project spanning Community Brain, Recap Flow, personal website, and Hermes.

📎 SHARED RESOURCES

Links and resources shared during the call:

Patchout Tech: https://patchoutech.com/en/ (shared by Patrick Chouinard in chat)

ntfy.sh: https://ntfy.sh/ (shared by Patrick Chouinard in chat)

🔄 FOLLOW-UPS WORTH EXPLORING

Paul Miller to send Brandon Hancock the link to Theo's T3Code video discussing CLI Proxy.

Ty Wells to look into installing/trying CLI Proxy as a possible replacement or supplement to his self-built subscription-switching system.

Patrick Chouinard to continue expanding his personal FinOps dashboard by adding GitHub (and other providers) over time.

Shakur Abdullah plans to write up and post about his AI "game of horse" model-comparison experiment.

Patrick Chouinard will try to get in touch with Brandon by the end of the week to confirm his availability, and will update the weekly post with the outcome.

1 comment

RecapFlow : October 13th Coaching call analysis