Thoughts on AI, building systems, and lessons from the trenches. https://karanbansal.in/blog/ en-us Fri, 04 Sep 2026 00:00:00 GMT https://karanbansal.in/blog/per-message-effort/ https://karanbansal.in/blog/per-message-effort/ Fri, 04 Sep 2026 00:00:00 GMT Claude Code stopped me with a warning that switching effort re-reads the full conversation. Fixed upstream three days earlier, shipped dark in the client, and switched on in 2.1.260 while I was writing this. The mechanism, the binary archaeology, and the day the dialog died. https://karanbansal.in/blog/banned-from-math/ https://karanbansal.in/blog/banned-from-math/ Thu, 13 Aug 2026 00:00:00 GMT An ITR-filing skill went from 37 to 628 GitHub stars off one stranger's tweet. The design rule that earned it: the LLM reads your documents and interviews you, but every rupee comes from tested Python. The one-in-350,000 rounding bug the fuzzer caught, and the statute bug it never could. https://karanbansal.in/blog/loop-engineering/ https://karanbansal.in/blog/loop-engineering/ Mon, 20 Jul 2026 00:00:00 GMT I read the first 182 replies to Gergely Orosz's 'can you share loops you regularly use?' One linked a loop you could actually run. Seven copy-paste loops in the order to learn them: /goal, /loop, the bounded Ralph, Stop hooks, dynamic workflows, cloud routines, and the full stack, with the termination checks, cost mechanics, and guardrails that keep them from burning $6,000 overnight. https://karanbansal.in/blog/good-morning-tax/ https://karanbansal.in/blog/good-morning-tax/ Wed, 08 Jul 2026 00:00:00 GMT My most expensive API requests ever are an acknowledgment, a menu choice, and a parade of good mornings. Receipts from 12,992 real Claude Code requests: the 60-minute cache cliff, the subagent meter, and why your bill tracks what the harness re-establishes, not what you type. https://karanbansal.in/blog/prompt-caching/ https://karanbansal.in/blog/prompt-caching/ Tue, 07 Jul 2026 00:00:00 GMT Prompt caching is the single biggest cost lever for LLM agents. But it doesn't cache your prompt. Here's what it actually caches, why prefixes must match exactly, and the rules that turn a $4.50 agent run into $0.62. https://karanbansal.in/blog/speed-up-ai-era/ https://karanbansal.in/blog/speed-up-ai-era/ Tue, 19 May 2026 00:00:00 GMT AI made optimizing your own workflow nearly free. The bottleneck moved from whether it's worth it to whether you even notice. Four levers (Connect, Encode, Teach, Parallelize), each with a tell you can catch yourself in. https://karanbansal.in/blog/claude-code-lsp/ https://karanbansal.in/blog/claude-code-lsp/ Sat, 28 Feb 2026 00:00:00 GMT Every Claude Code user is running without LSP. That means 30-60s grep searches instead of 50ms precise answers. Here's how to enable it: setup, real debug data, and undocumented discoveries. https://karanbansal.in/blog/claude-code-hooks/ https://karanbansal.in/blog/claude-code-hooks/ Sat, 25 Jan 2026 00:00:00 GMT Hooks are Claude Code's most underrated feature. From the event logger to safety hooks that block dangerous commands and protect secrets, plus a complete guide with ready-to-use code.