#Experiment

The Web Speech API + Gemini + VOICEVOX setup is complete — an AI character you can actually have a voice conversation with. Key implementation notes and impressions.

Implemented all 12 text processing tools planned in the previous article. Also reorganized the category system and switched the listing UI to a table layout.

Design patterns for reducing context usage and API calls in the AI auto-dev loop: blocking waits, read-forbidden files, and session isolation.

Setup notes for Qwen-Image-Edit-2511 on RunPod's RTX 4090 ($0.34/hr) using the ComfyUI template. Includes the fal Multiple-Angles LoRA (4 elevations × 8 azimuths × 3 distances) and a per-image cost breakdown that ends up cheaper than buying a 4090.

Running the Claude Code + Codex auto-loop for real. Generated 1,134 lines of game code.

Technical prep for automating an implement → review → fix loop with Claude Code and OpenAI Codex via tmux. Can it build something overnight unattended?

Comparing how to give AI characters personality across the Gemini, Claude, and OpenAI APIs — covering System Prompts, Prefill, and structured output.

How to restore a remote branch deleted with git push --delete, using GitHub API activity logs.

Dug up code from an audio preview player I built around 2015. Looking back at the struggle, and how simple it would be today.

An experiment in exchanging WebRTC signaling data via QR codes to achieve P2P voice calls with zero servers. Covers SDP chunking/reassembly and ICE candidate gathering.

A click-to-color mechanism using SVG paths. Explaining this surprisingly simple implementation — just changing the fill attribute — in vanilla JS.

A look back at a browser-based karaoke scoring tool I built years ago, and thoughts on how I'd approach it today.