AI Developer Accelerator — Coaching Call - May 12 VIEW RECORDING - 126 mins (No highlights) Meeting Purpose To share AI development progress, discuss challenges, and exchange best practices. Key Takeaways - Hybrid AI Workflows: Combining Claude Code for primary development with Codex for adversarial validation (e.g., security red-teaming) is a powerful pattern to counter model biases and improve code quality. - AI-Driven Infrastructure: AI can automate complex infrastructure tasks, like generating Disaster Recovery (DR) plans. The ideal customer is a high-growth startup needing enterprise-grade resilience, not a large enterprise. - User-Driven Development: A custom feedback widget enables users to narrate issues directly, capturing screen recordings and network logs. This provides high-fidelity input for AI agents to generate targeted fixes. - AI for Process Automation: AI can automate repetitive tasks by generating deterministic scripts (e.g., Google App Scripts) that run efficiently without repeated token consumption. Topics AI-Assisted Development Workflows - Hybrid Workflow (Claude Code + Codex): Rationale: Mitigates model biases by using two distinct models. Process: Claude Code generates specs, plans, and code. Codex performs adversarial validation and security red-teaming. Codex's analysis reports feed back into Claude Code for fixes. - Superpower Plugin: A recommended plugin for Claude Code that automates the SDLC (spec, plan, implement). It spawns sub-agents with fresh context windows to manage token limits and improve focus. - Skill Management: Recommendation: Build skills using Claude Code's skill creator skill, rather than downloading them. Rationale: Avoids supply chain security risks (e.g., prompt injection, poison pills). - UI Testing with Playwright: The Playwright MCP (Connector) allows Claude Code to execute UI tests on a local browser. Benefit: Enables automated, visual validation of UI output, which AI agents cannot perform directly.

📝 SUMMARY This week's call packed actionable intelligence on agentic development workflows, from dual-AI validation strategies to full-stack testing innovations. Patrick Chouinard shared his evolving "red team" approach using Claude Code alongside OpenAI Codex for adversarial validation, plus updates on a community RAG system and regulatory-driven SDLC requirements. Ty Wells demoed a user feedback hub and "Isotope" testing framework, while Morgan Cook showcased a digital signage engine and Elena presented a multi-cloud disaster recovery agent. The session wrapped with deep Q&A on Claude Code architecture, email automation safety guardrails, and targeting the right customer segment for AI consulting engagements. 💡 KEY INSIGHTS Red-team with different providers. Validate Claude Code output using OpenAI Codex (or vice versa) rather than the same model. Different biases mean flaws invisible to one may be caught by the other. Codex as adversarial harness. Prompt Codex to "destroy the application" when reviewing Claude Code output, then feed the failure reports back for remediation. Codex also includes GPT Image 2 in its $20/month subscription, useful for agentic image generation workflows without separate API costs. Security hygiene for AI skills. Avoid downloading skills from external repositories due to supply chain risks. Instead, use Claude Code's built-in skill creator to generate skills from completed work, or recreate them from reference. Agent design principles. Split agent responsibilities by context and single goals rather than human work logic. Giving one agent opposing objectives creates instability. It is acceptable for agents to have uneven workloads. Code generation vs artifact generation. When accuracy matters (like drawing precise graphics), ask AI to write code that generates the artifact rather than generating the artifact directly. AI evaluates code correctness better than visual correctness. Process before automation. AI amplifies broken processes. Clean up workflows before adding AI, then convert stable processes to deterministic scripts that run without AI for speed and reliability.

Turns out you can build a brain that remembers 68 coaching calls for the price of a fancy coffee—Patrick proved that data prep beats model selection when he spent only $0.75 of his $30 budget on cleaning versus embedding. Now he's packaging that Community Brain for download, and several of us are plotting our escape from Claude Code to Codex. If you missed the great migration debate, you’ll want a front-row seat this week. 📞 HOW THE CALLS WORK The calls can run 2+ hours. We want to make sure we're respecting everyone's time. Especially those of you who actually show up. Here's the structure: 👉 Reply to this post with your questions before the call 👉 If you submit a question and you're on the call, you go first 👉 We work through questions in the order they came in 👉 Then we open it up for everyone else If you can't make the call but want your question answered, drop it in the comments. We'll get to it. But priority goes to people who are there. The goal is simple: if you're taking the time to show up, you shouldn't have to wait behind questions from people who aren't even on the call. We've got some serious follow-ups brewing: Patrick is bringing the Community Brain download ready for anyone who wants to index their own chaos, he and Paul are reporting back on their Codex $100 tier test drives, and Morgan is handing over those Raspberry Pi kiosk scripts to Juan and Paul. If you were curious about any of those rabbit holes, now's the time to dig deeper. 🔗 ZOOM LINK (save this) https://us06web.zoom.us/j/81995207847?pwd=Xe6u6LmIQOmCP5VTnOwWYjDBfZNKGB.1 📅 WHEN Tuesday May 12th at 6PM ET Looking forward to seeing you on the call!

AI Developer Accelerator — Coaching Call - May 05 VIEW RECORDING - 71 mins (No highlights) Meeting Purpose Review AI projects and discuss developer tool strategies. Key Takeaways - Data Preprocessing is Key: Patrick's "Community Brain" RAG project shows that meticulous data preprocessing (costing ~$30) is the most valuable step, yielding an ~80% performance match to Opus with a local GPT-OSS model. - New AI Dev Stack: Superpower is emerging as a critical framework that enables a shift from Claude to Codex. This move is driven by Claude's recent performance issues and Codex's new, more generous $100/mo tier. - Hardware-Software Synergy: Morgan's Raspberry Pi kiosk engine provides a robust, offline-first display solution, which Juan plans to integrate with his "AI Booth Studio" for dynamic event content. Topics AI Developer Stack: Claude vs. Codex - Problem: Developers are experiencing performance degradation with Claude, including increased token usage and slower response times, leading to frustration and stalled progress. - Solution: Shifting to Codex, with Superpower as the common development harness. Superpower provides a consistent, high-quality development experience across platforms (GitHub Copilot, Claude Code, Codex). Codex's new $100/mo tier offers more generous usage limits than Claude's equivalent plan. - Transition Strategy: Patrick: Plans to switch from Claude Max to Codex Max next month. Paul: Will adopt Codex and Superpower to resolve a stalled project. Bastian: Shared methods for remote access to a local Codex instance (QR code or SSH).

📝 SUMMARY This week's roundtable featured Patrick Chouinard demonstrating his Community Brain RAG system that indexes 68 sessions of call transcripts, Juan Torres showcasing an AI photo-booth app for events, and Morgan Cook presenting a Raspberry Pi digital signage kiosk built over a weekend. The group also debated the shifting landscape between Claude Code and Codex, with several members planning migrations due to quality concerns and billing quirks. 💡 KEY INSIGHTS Data preparation beats model selection in RAG pipelines. Patrick found that cleaning and chunking transcripts cost only $0.75 of a $30 total spend, while embedding model choice (Nomic vs Gemini) showed minimal difference when underlying data was solid. Hybrid retrieval is essential for conversational data. Combining BM25 keyword search with vector search significantly outperformed either method alone when querying call transcripts. Small local models can reach 75-80% of frontier quality. With strong retrieval infrastructure, GPT-4o-mini and open-source alternatives approached Opus-level results at a fraction of the cost, though Gemma 4 struggled with complex retrieval logic despite its 30B parameters. Enterprise Claude Code requires configuration hardening. Deploy organizational settings.json files to supersede personal configs, and use seeded CLAUDE.md files to explain constraints rather than just blocking users. Watch for billing mode triggers. Mentioning competitor agent names like OpenClaw or Hermes in Claude Code sessions switches from subscription to API token billing. The Superpower framework now bridges Claude Code, GitHub Copilot, and Codex. This allows workflow portability as members evaluate switching between coding agents. Model default preferences are the new SEO. If a technology isn't in a model's training-weighted defaults, it effectively doesn't exist for AI-assisted development. ❓ KEY Q&A Q: What replaced ShipKit for the group? A: Most have moved to Claude Superpowers, which now works across Claude Code, GitHub Copilot, and Codex.