An Hour Down Claude Code's Memory Hole
Claude Code's auto-memory feature ships on by default and eats 47% of every system prompt. I spent an hour debugging the symptoms before finding a one-line fix.
Read more →I Sent the Same Prompt Injection to Ten LLMs. Three Complied.
I sent the same five-line XML message to ten commercial LLMs. Seven ignored it. Three dumped everything I planted as structured JSON.
Read more →Your AI Reviewer Has the Same Blind Spots You Do
We sent one plan to five AI model families for independent review. They found seven issues we missed, including a regex that would crash the build.
Read more →We Searched the Agent Skills Ecosystem for SEO
We searched every Agent Skills repository for SEO coverage: official, community, broad GitHub. Zero results. Then we found the gap was bigger than SEO.
Read more →Our AI Teams Had a Communication Problem (The Fix Was From 1995)
We built three AI teams that worked great alone but couldn't coordinate. The user became the message bus, and the message bus forgets. The fix was a filesystem protocol designed by Daniel J. Bernstein in 1995.
Read more →Our AI Critic Was Going Easy on Us (Research Told Us Why)
We run a team of AI personas that collaborate in shared context. One day we asked: is our critic actually critiquing, or is the same model rating its own work? Seven research papers and one real test later, we built a feature to fix it.
Read more →Building a Claude Traffic Proxy in One Session
I wanted to track how much my Claude API usage was actually costing me. Per request. Per task. Per tool call. So I built Langley: an intercepting proxy that captures every Claude API request, extracts token usage, calculates costs, and shows it all in real-time.
Read more →Why I Make Claude Argue With Itself Before Writing Code
I asked Claude to "make my scraper robust." It generated 200 lines of plausible-looking code. All garbage. Here's what I do instead.
Read more →