Weekly Coaching Call Recording 8/5/2025 · AI Developer Accelerator

Weekly Coaching Call Recording 8/5/2025

View Recording Here:

Key Takeaways

Rising Costs are Forcing Tool Changes: Developers are actively seeking alternatives to platforms like Cursor and Claude Code due to new rate limits and rapidly increasing costs, with a $20/month budget lasting only a few days for some users.
No-Code/Low-Code for Prototyping is Key: Tools like N8n are being widely adopted for creating quick proofs-of-concept (POCs) to demonstrate value to clients, secure buy-in, and model workflows before committing to complex coding.
The "AI Bubble" Concern: There's a growing discussion about the long-term financial viability of AI development, with concerns that current model usage is heavily subsidized by providers and that costs may become unsustainable for many users and businesses.
Open-Source & Local Models are Gaining Traction: As proprietary model costs rise, there's renewed interest in open-source models (like OpenAI's new release) and local development, especially with hardware advancements on the horizon.
Strategy Over Brute Force: Experienced users are emphasizing sophisticated workflows, such as detailed planning phases (Andrew), multi-tool strategies to minimize costs (Patrick), and robust evaluation (Eval) processes to manage projects effectively.

Topics

AI Development Tools & Platforms

Cursor & Claude Code: Users report burning through their $20 monthly credits in just a few days due to recent changes. The lack of clarity in usage dashboards is a common frustration.
N8n (No-Code): Praised as an excellent tool for building POCs, modeling workflows, and getting client buy-in. However, users caution that complex flows with many nodes can become difficult to debug.
Google Opal: Found to be good for generating initial one-shot prompts and flows, but users struggle with editing and iterating on the generated workflows. It also uses all data for training.
LlamaIndex & LangChain: LlamaIndex is preferred for RAG-specific tasks due to its extensive toolkit, while LangChain is seen as more general-purpose but complex.
Convex: Discussed as a powerful backend-as-a-service with built-in RAG capabilities, appreciated for its speed and unopinionated nature.
Emerging Alternatives: The group is exploring tools like Kiro (for its detailed planning output), Gemini CLI, Open Code (a console-based version of Cursor), and Warp to circumvent the costs of mainstream platforms.

AI Models & Performance

New OpenAI Open-Source Model: Early impressions are positive. It's considered impressive for its performance on older hardware (GTX 1080 Ti) and its small memory footprint.
Claude Opus 4.1: Perceived as an incremental improvement, with minor performance gains across most tasks.
Model Costs & The "AI Bubble": A major theme was the concern that AI companies are heavily subsidizing usage and that the current "free lunch" will end, potentially leading to a market correction or a shift to more efficient, local models.
Local & Open-Weight Models: Models like Kimi K2 and GLM 4.5 Air are showing promise, capable of impressive feats like one-shot coding a Space Invaders game on consumer hardware, signaling a potential shift towards local execution.

Development Strategies & Workflows

The Planning Phase: Andrew shared a detailed strategy of using "plan mode" extensively in Claude Code, iterating on a design document and implementation plan with the AI before writing any code to save tokens and avoid errors.
Prompt Engineering: A new back-and-forth prompting method was shared that helps refine a user's intent and pull out ambiguities before generating the final output.
Evaluations (Evals): Evals are seen as the "secret sauce" and a critical, though difficult, part of development. Strategies include using an LLM-as-a-judge, manual spot-checking, and making the client responsible for validating the output.
Cost-Saving Multi-Tool Approach: Patrick described a workflow that uses cheaper or free tools for different stages: GitHub Copilot and Gemini CLI for documentation and planning, and saving expensive Claude Code tokens for the final execution and debugging steps.

Client Acquisition & Project Management

Finding Clients: Participants are networking at events and tapping into personal networks (property management, digital marketing, HVAC) to find opportunities.
Demonstrating Value: Using N8n to quickly build a POC on a call with a client is a powerful sales tool that helps them visualize the solution and gives them a sense of ownership.
Managing Scope Creep: A significant challenge is that clients often believe "anything is possible" with AI. Setting realistic expectations and nailing down the project scope is crucial to success.
Finding a Niche: Getting into an industry vertical, like insurance, and identifying common problems can lead to repeatable, valuable solutions.

Current AI Projects

Andrew: Building a graph RAG project using LlamaIndex and the Kuzu embedded graph database.
Al: Automating business processes for property management, digital marketing, and HVAC services, starting with N8n for workflow modeling.
Abdel: Using Google Opal to create an Ideal Customer Profile generator for a Korean client in the environmental/pharmaceutical industry.
Abhi: Creating a POC with Google ADK to build an agent that automates L1/L2 technical support tasks, integrated with ServiceNow and Jira.
Mark: Building an app that consumes a YouTube transcript MCP in Docker to display the latest videos from selected influencers.
Mitch: Maintaining a video generation workflow that uses Gemini 2.5 Pro and a Supabase backend to create clips based on a master prompt.
Patrick: Using ChatGPT Agent Mode to read, categorize, and generate summary responses for posts on the School community board.
Neil: Working on a data drift detection portfolio project using agents for investigation and strategy creation with Google ADK.
Stephen: Getting started with agentic frameworks to build an observability platform for correlation and root cause analysis.

Action Items

Al: Use Apify to scrape Google reviews for property management companies to identify customer pain points (0:37:17).
Abdel: Post in the School community to find Americans in the environmental or pharmaceutical fields for his Korean client (0:53:37).
Mark: Research N8n as a tool for building proof-of-concept workflows before coding (0:58:34).
Neil: Consider using Lovable for building a more polished UI for his data science portfolio project instead of just Streamlit (1:49:03).
Prem: Look into using Next.js with Cursor for building a UI, as it can generate impressive results with few prompts (1:52:33).
Prem: Investigate "Repo-Prompt" as a tool for better context management and searching within a codebase (1:57:02).

P.S.

Thank you to

for hosting this weeks call! You are the man!

We are on vacation right now and I want to let you know that we survived the night dive and we got to swim with a ton of sharks!

0 comments