1d • General discussion
AI Developer Accelerator — Coaching Call - June 16th
AI Developer Accelerator — Coaching Call - June 16
Meeting Purpose
Key Takeaways
- Fable's absence is a setback, but developers are adapting with multi-model workflows (e.g., Claude for planning, Codex for coding) and focusing on deterministic scaffolding to maintain control.
- New projects include an AI-native CRM for UK estate agents (Ryan), a positive-only "KindMark" platform for service workers (Ty), and an AI photo booth (Juan).
- Patrick developed "AgentOps," a meta-scaffolding for Hermes using NATS as a "nervous system" to monitor and manage his home lab infrastructure.
- A key strategy is using an "adversarial" system prompt to force the AI to challenge assumptions and clarify the core business problem, preventing it from amplifying broken processes.
Topics
Fable's Absence & Mitigation Strategies
- Fable's shutdown is a significant setback, but developers are adapting with multi-model workflows.
- Ryan: Uses Opus for iteration on a Fable-generated V1 plan.
- Ty: Used Fable for initial problem-solving but found it unusable for security auditing, a critical step.
- Paul: Combines Opus for projects with Codex for validating requirements documents.
- Patrick: Prefers a collaborative model that challenges assumptions over one that simply executes instructions.
New Projects & Development
- Ryan's AI-Native Estate Agent CRM Goal: A complete AI-native operating system for UK estate agents. Features: Aggregates communication (email, WhatsApp, property portals). AI assistant drafts responses and learns from user edits. Integrates with prop-tech APIs for automated marketing (e.g., sending letters to neighbors after a sale). RAG System: Chunks communication history and meeting transcripts. Chunking Method: Uses an N8N workflow to restructure transcripts into linear, meaningful blocks before traditional chunking. This ensures questions and their answers are kept together, improving retrieval quality.
- Ty's "KindMark" Platform Goal: A positive-only platform for recognizing service workers' character traits (e.g., patience, kindness). Mechanism: Users submit "kind marks" that are privacy-first (PII removed). Workers can claim these marks to build a character profile. Data Ingestion: Scrapes existing reviews (Yelp, Google) to identify and frame positive attributes.
- Patrick's "AgentOps" Goal: A meta-scaffolding for Hermes to create an "operator environment" for his home lab. Architecture: NATS: A messaging system acting as a "nervous system" to standardize and centralize messages from monitoring tools (Uptime Kuma, Prometheus, Grafana). Hermes: Consumes NATS messages and uses a toolbox of scripts to intervene on infrastructure (Proxmox, router). Security: Uses Authentic for authentication and In Physical for secret management. Development: Built with Codex 5.5, which Patrick finds superior to Claude for terminal and infrastructure code.
- Juan's AI Photo Booth Goal: An AI photo booth application using diffusion models. Workflow: Uses Wavespeed's desktop app and API to experiment with models and LORAs (Low-Rank Adaptations) without owning GPUs. Model Selection: Found Chinese models (e.g., Quent-Image-2, Hoonjuand-Image-3, Seadream-4.5) to be highly dynamic. FinOps: The app will track API costs per event to determine client pricing.
AI as a Personal Assistant
- Patrick used Claude to manage the administrative tasks after his mother's passing.
- Process: Scanned funeral home documents → Claude created a timeline, to-do list, and identified necessary government forms and programs.
- Result: Reduced a two-month process to 48 hours, demonstrating AI's value in high-stress, low-bandwidth situations.
Best Practices & Methodologies
- Adversarial System Prompt Problem: Clients often state a desired solution instead of the core business problem, leading to AI amplifying broken processes. Solution: Patrick uses a system prompt that makes the AI act as a "devil's advocate," challenging assumptions and forcing clarification of the underlying problem. Rationale: This mirrors the role of a good business analyst and prevents building the wrong solution.
- Ty's "Intent Queue" Problem: Using the "by the way" command in Claude Code doesn't save tokens, as the full context is still processed. Solution: A local queue system to save feature ideas for later. Mechanism: Use "by the way" to capture an idea. The AI's response is saved to a local file. Clear the session. A hook in the next session prompts to process the queued items. Benefit: Saves tokens by re-priming the context only once per feature, not for every "by the way" query.
Next Steps
- Ryan: Continue iterating on the estate agent CRM; aim for a usable version in ~2 weeks.
- Ty: Publish the "Intent Queue" concept plan to the Skool community.
- Patrick: Continue developing "AgentOps"; share when it is secure and stable.
- Juan: Meet with a pilot partner tomorrow; aim for a pilot go-live this weekend or next.
- Paul: Test the "adversarial" system prompt with a client.
Action Items
- Review 20.com CRM for reuse/integration - WATCH (5 secs)
- Post 'by the way' intent queue concept in Skool for Paul, Patrick, Juan, Ryan, Morgan - WATCH (5 secs)
0
0 comments
AI Developer Accelerator — Coaching Call - June 16th
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
+6
2
+3
3
+2
4
+2
5
+1
Powered by