Skip to content

Loop Notes

  • AI Coding Agents
  • Backend & Servers
  • Tech in Korea
  • Hiring & Careers

jjeongil

I'm a backend developer in Seoul. I use AI coding agents such as Claude Code in my daily work, and I write about them and about the Korean apps and services visitors get stuck on.

Claude Code Auto Mode Not Working: Check the Real Mode First

November 18, 2026 by jjeongil
Diagram of a terminal status line reading "manual mode on" with arrows to four checks: another mode name, unavailable for this model, disabled by settings, and claude -p permissionMode

Claude Code auto mode not working? Cycle Shift+Tab once: if auto is in the cycle, a defaultMode line wins; if not, the text beside it names the cause.

Categories AI Coding Agents

Ollama Not Using GPU After an Update on Linux: The Fix

November 17, 2026 by jjeongil
Simple diagram of Ollama going from 100% CPU to 100% GPU after a reboot on an Ubuntu machine with an NVIDIA driver mismatch.

Ollama on 100% CPU after an Ubuntu update? A driver mismatch: new NVIDIA packages, old module still loaded. Reboot, check ollama ps, then hold the driver.

Categories AI Coding Agents

Claude Code Ignores CLAUDE.md or Skills? Where to Look First

November 16, 2026 by jjeongil
Diagram of four Claude Code setup layers, CLAUDE.md, subfolder rule, skill and hook, each checked for whether it loaded and was called

If Claude Code ignores CLAUDE.md, check what loaded: every loaded rule held 45 of 45 in our runs; misses were subfolder rules and uncalled skills.

Categories AI Coding Agents

Local LLM Hardware: What to Buy for the Model You’ll Run

November 15, 2026 by jjeongil
Diagram of a model block made of model weights and context stacked together, set beside a 16GB card, a 24GB card and a 96GB Mac drawn as containers of increasing size

Local LLM hardware: a 14B fits a 16GB card at 32K, a 30B MoE coder at 32K takes 24GB, and 70B at 32K needs a 96GB Mac, because context uses memory.

Categories AI Coding Agents

Claude Code Skill Not Triggering? It’s the Prompt (45 Runs)

November 14, 2026 by jjeongil
A request that names the job shows a Skill line; one that talks around it shows none, so call /skill-name

If a skill shows in /skills but Claude ignores it, the description likely isn’t why: in 45 runs, misses came only from requests that never named the job.

Categories AI Coding Agents

1Password for Claude: My Mac Setup and What It Can’t Do

November 13, 2026 by jjeongil
1Password connector shown as connected in Claude settings

1Password for Claude fills logins on a Mac after you approve each one. What it needs, how it asks at every login, and what to try when it stops connecting.

Categories AI Coding Agents

Coding Agent Model Routing: When a Cheap Model Is Enough

November 14, 2026November 5, 2026 by jjeongil
Diagram routing coding agent tasks by risk: Low goes to a cheap model, Medium starts cheap and moves to a strong model after one failure, High goes to a strong model and then a person approves

Send coding-agent tasks to a cheap model like GPT-6 Luna only when you can check the result fast, and route the rest by what a wrong change would break.

Categories AI Coding Agents

ChatGPT Pro 500 vs Pro 200 for Codex: Why I Went Back

November 4, 2026 by jjeongil
Two timelines of the same agent task: model generation blocks get shorter with a faster model, while test and build waits stay the same, so total time shrinks only a little

Pro 500 vs Pro 200 for Codex after the October 30 cut: what each plan gives, how Ultrafast burns 8x usage, and why I moved back to Pro 200.

Categories AI Coding Agents

Letting Claude Code Run Overnight: What I Leave Open

November 16, 2026November 3, 2026 by jjeongil
Diagram sorting coding agent actions into open, blocked and not there by one test: can I undo it

Before Claude Code runs unattended overnight, I keep push keys and real .env files out of its reach. Then I don’t trust “success”: 38 blocked runs said it.

Categories AI Coding Agents

AI Agent Access to Email and Payments: Read, Recommend, Act

November 15, 2026November 2, 2026 by jjeongil
Diagram of three stages, read, recommend and act, with a stop before act marked you approve

Let an AI agent read your email and draft replies, but approve each send and payment yourself and leave off the one setting that removes that stop.

Categories AI Coding Agents
Older posts
Page1 Page2 … Page5 Next →

Recent Posts

  • Claude Code Auto Mode Not Working: Check the Real Mode First
  • Ollama Not Using GPU After an Update on Linux: The Fix
  • Claude Code Ignores CLAUDE.md or Skills? Where to Look First
  • Local LLM Hardware: What to Buy for the Model You’ll Run
  • Claude Code Skill Not Triggering? It’s the Prompt (45 Runs)
  • About
  • Privacy Policy
  • Terms of Use
  • Editorial Policy
  • Contact
© 2026 Loop Notes