← back
Vibe Engineering Effect Apps — Michael Arnaldi, Effectful
Takeaway
To make coding agents useful on undocumented libraries, drop the library repo straight into your project so the agent's code-focused training kicks in.
Summary
- Live unscripted workshop building an Effect-TS project from scratch with GPT-5 in agentic mode — author hasn't hand-written code in ~6–8 months.
- Key trick for library-level coding: clone the library's repo directly into your working directory so coding agents (post-trained to focus on user code, ignore node_modules/gitignored files) treat it as first-class context.
- Frames LLMs as stateless after pre-training: 1M-token context windows can hurt by confusing the model on multi-task contexts.
- Stack target: bun + Vitest + Effect HTTP server with OpenAPI, type-safe client, workflow clustering for persistent ops.
- Compares GPT-4 ('kid running through the house with a knife', per Geoffrey Huntley's Ralph loops) vs GPT-5 reliability.
effectcoding-agentstypescript
Original description
What if the best way to get coding agents to use a library well is not better prompts, but giving them the library's actual code? In this workshop, Michael Arnaldi walks through a practical approach to building with Effect and LLMs by cloning the Effect repo into the project, extracting patterns directly from the source, and using those patterns to guide agent behavior. Starting from an empty repository, the session shows how to set up an Effect-based app with tests, strict TypeScript diagnostics, agent instructions, and a simple HTTP API, while also exploring the broader problem of how to make agents effective in unfamiliar codebases. If you're building with coding agents and care about reliability, structure, and real-world Effect workflows, this is a useful hands-on framing. Speaker info: - https://x.com/MichaelArnaldi - https://www.linkedin.com/in/michael-arnaldi-52858114a/ Timestamps 0:15 – Introduction and context setting for the workshop 0:47 – Interactive audience poll on experience with Effect and AI tooling 3:16 – Discussing the core philosophy: "Just clone the repo" for AI context 5:59 – Understanding LLMs vs. the human brain and context window limitations 13:13 – Project setup: Starting from an empty repository 14:20 – Initializing the project with Bun, Vitest, and TypeScript 19:18 – Adding Effect beta and configuring TSGo for the compiler 30:30 – Configuring strict diagnostics for AI-assisted development 35:20 – Adding the Effect repository as a git subtree for better agent access 37:07 – Creating agents.md to establish rules and available commands 41:40 – Researching Effect patterns for building an HTTP API 43:08 – Discussing "Spec-Driven Development" and avoiding plan-mode limitations 54:02 – Drafting the plan for the Todo HTTP API 1:05:07 – Implementing the SQL client and migration patterns 1:13:42 – Reviewing API schemas and handling identified code duplication 1:18:14 – Starting the API server and verifying OpenAPI documentation 1:20:56 – Cleaning up test suites and enforcing best practices for layers 1:38:08 – Concluding remarks on workflows, clustering, and future stability in Effect