From the Blog

Thoughts on Open Source, AI, web development, and building software

Found 110 posts

2026-06-18

My agent kept losing facts it had already saved

AIdaemon could save a fact and then fail to find it when I asked for it. Semantic search ranks by what's on the same topic, so the short fact holding the real answer gets buried under wordier ones nearby. A reranker reads each candidate against the question and puts the right one back on top.

aisoftware-developmentopen-source

2026-06-08

13 min read

Local Gemma was too slow with AIdaemon until I fixed llama.cpp and the prompt size

I wanted AIdaemon on local Gemma 4 26B through llama.cpp, not Ollama. Generation ran at ~45 tok/s on an M4 Pro. Agent turns still felt stuck because prefill on 14k-token prompts took 8 to 9 seconds before the model wrote a single word.

aisoftware-development+1

2026-05-20

12 min read

How production-grade AI agents actually work

A plain-English walkthrough of the control plane, tool plane, memory, guardrails, sub-agents, and observability behind production-grade AI agents.

AIArchitecture+1

2026-05-15

3 min read

How to work with any generative AI

Switching from ChatGPT to Claude to Gemini to Grok, or to any open source model, shouldn't mean relearning how to prompt, what to trust, and what's safe to hand off. Under the hood they share the same handful of traits, and the same few habits get good results from all of them.

aifundamentals

2026-05-14

2 min read

Skills, CLAUDE.md, subagents, hooks, or MCP, when to use which

Claude Code has five customization layers. Pick the wrong one and you get a Skill that never fires, a hook doing subagent work, or a CLAUDE.md so bloated it eats half your context before you type. Here's where each one belongs.

aiclaude-code+1

2026-05-11

5 min read

Cutting OpenNext KV Storage Costs on Cloudflare Workers

OpenNext on Cloudflare stores ISR cache in KV, and every deploy quietly adds a new universe of keys that never get cleaned up. Here is how one of my sites went from $0.50/day in KV storage to under three cents.

cloudflarenextjs+1

2026-05-11

3 min read

Notes on installing Microsoft Clarity

I added Microsoft Clarity to one of my projects this week. It is free, the setup took ten minutes, and it has already told me more about how people use the site than Google Analytics ever has.

web-developmentanalytics+1

2026-05-03

5 min read

How to keep multiple coding agents from overwriting each other

The practical workflow I use for parallel AI coding agents: one worktree per agent, one branch per task, clear file ownership, protected main, and shared instructions through AGENTS.md or CLAUDE.md.

aisoftware-development+1

2026-04-29

5 min read

Choosing Between CLI and MCP for Your AI Agent Workflows

After a year of building agent workflows, I've learned when CLI makes sense versus when MCP is worth the investment. Here's my decision framework.

ai-toolsdevelopment-tools+1

2026-04-25

4 min read

How OpenNext's regional cache reduces Workers CPU on every cache hit

withRegionalCache is a one-line wrapper that ships with @opennextjs/cloudflare. It puts a Workers Cache API layer in front of your KV cache and lets you skip the D1 tag-cache lookup on hits. For content sites, that's the difference between burning real CPU on every request and barely doing any work at all.

web-developmentnextjs+1

Get the latest posts and insights delivered to your inbox.

No spam. Unsubscribe at any time.

From the Blog

My agent kept losing facts it had already saved

Local Gemma was too slow with AIdaemon until I fixed llama.cpp and the prompt size

How production-grade AI agents actually work

How to work with any generative AI

Skills, CLAUDE.md, subagents, hooks, or MCP, when to use which

Cutting OpenNext KV Storage Costs on Cloudflare Workers

Notes on installing Microsoft Clarity

How to keep multiple coding agents from overwriting each other

Choosing Between CLI and MCP for Your AI Agent Workflows

How OpenNext's regional cache reduces Workers CPU on every cache hit

Stay Updated