Writing
Technical breakdowns, practical guides, essays and reflections on building with AI.
Chat Windows Don't Teach. So I Built an AI That Checks.
Two 2025 trials show that asking a chatbot for answers raises practice scores and lowers exam scores. The Learning Harness is my fix: lessons in prerequisite order, a quiz the same day, an explain-back after a night of sleep, and Anki cards only for what passed.
The Autoresearch Loop
A visual guide to the propose, run, measure and decide loop, with the minimum parameters needed to make it work.
What I Call the Post-AI-Psychosis Depression
A reflection on the quiet crash that follows a few weeks of all-consuming momentum with an AI.
Reverse Engineering sarvam-code
A static-analysis teardown of Sarvam's coding-agent CLI: the Codex lineage, its verbatim system prompt, 24 tools, 50 feature flags, and what is India-specific inside the binary.
AskUserQuestion Belongs in Every AI's System Prompt
Why bounded-choice question widgets beat prose Q&A in AI conversations, with a custom instruction you can paste into ChatGPT or Claude.
Auto-Close XML Tags in Your Terminal for Better Claude Code Prompts
Set up automatic XML closing tags so you can write structured Claude Code prompts without the friction of typing them manually. macOS, one keystroke.
Get Claude Code Notifications on Your iPhone & Apple Watch in 1 Minute
Set up push notifications so your iPhone (and Apple Watch) buzzes the moment Claude Code finishes. Free, no server required.
Why You Can't Gradient Descent Your Way to Funny
The distribution width problem. Why scaling laws hit a wall at comedy while breezing past beauty.