Mar 25 • General discussion
RecapFlow : March 24th Coaching call analysis
📝 SUMMARY
A dense, high-signal call covering self-improving AI pipelines, governance-first agent architecture, Stripe best practices, biometric authentication, and mobile ideation workflows. The strongest through-line: the shift from using AI interactively to building systems that run autonomously — defining quality rubrics, letting agents evaluate and improve their own outputs, and removing the human from the loop wherever possible. Practical tool recommendations (CMux, Codex for autonomous tasks, Terraform for infrastructure, Discord over Telegram for agent memory) were grounded in real production experience. The IronClaw white paper, Ty's FaceGate SDK demo, and Patrick's RecapFlow auto-research experiment are the most concrete follow-ups to watch for in the coming week.
💡 KEY INSIGHTS
Self-Improving AI Pipelines — The Most Actionable Framework Shared This Call
Build systems that eliminate the human from the evaluation loop. Define explicit pass/fail criteria and a point-based rubric, build a representative input suite (e.g., 60 test cases), let the AI run experiments, grade its own outputs, identify failure modes, update its own system prompt, and iterate. Apply this at the individual pipeline step level first, then at the full system level. Brandon uses Codex for this because it runs autonomously for long periods without prompting for human confirmation. Expensive but produces measurable, compounding improvement.
The Hardest Part Is Defining "Good"
For mathematical outputs, scoring is straightforward. For language outputs, defining quality is the core challenge. Patrick's approach: use mechanical checks (did all URLs get extracted? is compression within bounds?) for the fast inner loop, and community feedback as the slow outer loop for subjective quality.
Governance Before Features for AI Agents
Prioritize governance before adding capabilities. Recommended architecture: read-only access to most systems, human-in-the-loop via Discord or Telegram for any state-changing action, full audit trail, and a smart router using local models (Ollama) for routine tasks and frontier models only for complex ones. This is the core principle behind the IronClaw framework.
Infrastructure as Code Is Underused by Vibe Coders
Terraform gives Claude Code a source of truth for cloud infrastructure state, makes changes reproducible, and eliminates the "what is the current state of my database?" problem. Claude Code knows Terraform well. If you've built infrastructure manually, ask Claude Code to inspect your existing setup and generate the Terraform files from it.
Building in Public Creates Inbound Opportunities You Cannot Manufacture
Patrick published a project to a public GitHub repo. An architect at a client company found it through GitHub's search agent and it led to a production contract. Public repos are a discovery surface that outbound outreach cannot replicate.
Every Business Runs on About 12 SOPs That Aren't Written Down
Every employee at a small business executes roughly 12 standard operating procedures that exist only in their head. The job of an AI consultant is to surface those SOPs, codify them, and build agents to execute them. You discover what context is missing by running the agent at scale and observing where it fails.
❓ KEY Q&A
Q: Best practices for implementing Stripe without getting lost in key management?
Store only the Stripe customer ID in your app database. Query Stripe directly at runtime for everything else — subscription status, plan details, payment history. This prevents stale data and simplifies your data model. Use the Stripe CLI (controlled by Claude Code or Codex) to automate product and key setup. Lemon Squeezy is a simpler alternative. For donations, PayPal has the best feature support.
Q: Are ShipKit projects still relevant?
Core concepts (chat, RAG, Vercel AI SDK) remain valid. The main shift is tooling: the community has largely moved from Cursor to Claude Code. CMux is now recommended for managing multiple instances. ShipKit is still the right starting point for beginners; advanced users should layer in newer tools on top.
Q: Best model for non-thinking consumer app use cases?
GPT-5.4 with thinking disabled has outperformed all other models tested for consumer-facing apps. Gemini Flash is the fastest option. For Claude Code specifically, Opus with the 1-million-token context window is recommended — the larger context reduces compaction interruptions during long sessions.
Q: Telegram vs. Discord for AI agent communication?
Both work. Telegram is easier to set up. Discord is better for complex setups because separate channels prevent context from different domains from polluting each other, allowing significantly more total content across agent memory. Patrick recommends Discord for any setup growing in scope.
Q: What is Databricks and when should you use it?
A lakehouse platform combining data warehouse speed with data lake flexibility. Handles analytical queries, raw data processing, ML, AI pipelines, and dashboards. Sits on top of GCP, AWS, or Azure storage using Delta Lake format. Best suited for data-heavy processing workloads.
Q: How do you handle ideation on the go?
Brandon: Dictate into Claude on mobile, have Claude Code save it as a task template to a branch, pull the branch at the desktop. Whisper Flow on iPhone recommended for fast voice-to-text.
Patrick: Uses ChatGPT voice (best voice model available) for extended ideation while driving, then asks it to compile a PRD. The PRD goes into CoworkOS, which breaks it into feature-sized tasks for Claude Code to execute.
Key friction point: ChatGPT has the best voice model; Claude has the best development model. They don't currently interoperate.
Q: Is 99.7% face recognition accuracy sufficient for authentication?
For time-and-attendance and shared-device use cases, yes — more secure than passwords. The system stores only an encrypted mathematical vector, not the image. A new vector is captured at every authentication event, effectively rotating the credential daily.
Important legal caveat: In the U.S., law enforcement can compel face unlock without a warrant but cannot compel disclosure of a passcode. Material consideration for any sensitive-data application.
🛠️ TOOLS AND CONCEPTS MENTIONED
CMux — Manages multiple simultaneous Claude Code instances. Significant productivity upgrade for power users. cmux.com
Claude Code with Opus and 1M context — Current recommended setup. Larger context window reduces compaction interruptions.
Codex (OpenAI) — Preferred over Claude Code for long-running autonomous tasks. Requires fewer human check-ins, better suited for multi-hour self-improving loops.
Whisper Flow — iPhone app for fast voice-to-text. Used to dictate ideas into Claude Code on mobile.
Terraform — Infrastructure as code. Gives AI agents a reliable source of truth for cloud infrastructure state.
Databricks / Delta Lake — Lakehouse platform for data-heavy AI pipeline deployments.
IronClaw — Secure, auditable, locally-hosted AI agent architecture with governance-first design and human-in-the-loop controls.
Ollama — Local model runner. Used in IronClaw to route routine tasks away from expensive frontier models.
Fieldy — Wearable ambient recording device. Transcribes accurately even in adjacent rooms. Offers a webhook for piping output to N8N or other automation tools.
Plaud — Wearable recorder that attaches to the back of a phone. Brandon's current recommendation when Fathom is blocked. Also records system audio via desktop app.
Limitless — Previously a leading ambient recording platform. Acquired by Meta. Existing users have until September 2026 before the service ends.
Eleven Labs — Text-to-speech platform. Scott uses it with a custom Jarvis voice for his daily AI news digest.
Vanta — Compliance automation platform. Ty is considering it for security review of his FaceGate biometric SDK.
CoworkOS (Patrick) — Patrick's configured Claude environment with memory, personality, and skills. Parses PRDs and hands off structured feature packages to Claude Code.
FaceGate (Ty Wells) — In-development face authentication SDK for web apps. Stores only encrypted face vectors, not images. Designed for shared-device environments.
Immersed — VR app for Meta Quest 3 that creates a multi-monitor coding environment.
Notion AI with MCP — Elijah runs an MCP integration to pull Notion context directly into Claude Code.
Session Buddy — Browser extension for saving and annotating large numbers of open tabs.
Lemon Squeezy — Stripe alternative. Simpler to set up but takes a higher cut.
📎 SHARED RESOURCES
Claude Certified Architect Foundations Exam Guide (Official PDF, 40 pages)
Claude Architect Certification Interactive Study App (flashcards, quizzes, video summaries, Q&A chatbot — built by Ty Wells)
Claude Certified Architect Exam Registration
CMux — Multi-Instance Claude Code Manager
Patrick's AI Personality Kit for CLI tools (Claude Code, Gemini CLI, Copilot CLI, Codex CLI)
Patrick's CMux Plugin for Claude Code
OpenClaw Security Architecture Guide (governance-first AI agent white paper)
NVIDIA NemoClaw (alternative secure Claude architecture for self-hosted AI)
Three Buckets AI Framework (Elijah Stambaugh — AI consulting and enterprise transformation)
Databricks Medallion Architecture Reference
Session Buddy (browser tab manager)
Accelerando by Charles Stross (Brandon's top book rec — audiobook preferred)
The Phoenix Project (systems thinking and SOPs — recommended for AI consultants)
🔄 FOLLOW-UPS WORTH EXPLORING
RecapFlow community experiment — Patrick is asking community members to comment on the weekly recap post. Comments will be fed into the auto-research improvement loop as training signal. The more specific your feedback, the better.
Ty's FaceGate SDK — Ty plans to have a testable version ready for the next call where community members can attempt to enroll, verify, and try to break the system.
Scott's IronClaw white paper — Scott plans to share the full 40-plus page document in the community. Recommended reading even if you never intend to implement it — the governance-first thinking is broadly applicable.
Scott's AI News Digest — Daily AI news email with an Eleven Labs Jarvis voice audio version.
Brandon's customer success manager agent — Brandon expects to have it built by the next call and will share results.
Elijah's competition win — Details embargoed for another week or two. Follow up at the next call.
Anthropic ToS for Claude subscriptions via Agent SDK — Using your Claude subscription via the Agent SDK in a web app may violate Anthropic's terms of service. Codex does not have this restriction. Verify your compliance posture before shipping.
Elijah's embed and equip consulting programs — Two offerings for legacy businesses adopting AI. Cohort forming now.
4
3 comments
RecapFlow : March 24th 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
+8
2
+3
3
+3
4
+2
5
+2
Powered by