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.

DGX Spark vs RTX 4090 vs Mac mini: Why I Kept My 4090

November 1, 2026October 19, 2026 by jjeongil
Two-step diagram for local LLM hardware: first check whether the model and context fit in memory, then bandwidth decides speed once it fits

DGX Spark: 128GB at 273GB/s. RTX 4090: 24GB at about 1TB/s. My measured 4090 VRAM numbers show when capacity matters, and why I haven’t bought a Spark.

Categories AI Coding Agents

Claude Code Custom Commands to Skills: Why /project: Fails

November 1, 2026October 18, 2026 by jjeongil
Diagram of a single command file in .claude/commands and a skill folder with SKILL.md in .claude/skills, both called as /deploy

/project: fails because Claude Code merged custom commands into skills. What changed, what broke when I moved mine, and the description fix that helped.

Categories AI Coding Agents

AI Agent Ran git commit –no-verify? Make CI the Real Gate

October 31, 2026October 17, 2026 by jjeongil
Flowchart: an AI agent's commit fails the pre-commit hook, is re-committed with git commit --no-verify, then fails lint in GitHub Actions CI

git commit –no-verify skips pre-commit and commit-msg hooks, so no hook can stop it. My AI agent used it; CI as a required check caught the lint error.

Categories AI Coding Agents

Claude Outage Fallback: Request Failed, Files Still Changed

November 1, 2026October 15, 2026 by jjeongil
Recreated Claude status incident record for September 29, 2026, showing the title Elevated errors and the 14:00–14:59 UTC impact window, with update rows hidden

Claude down mid-task? A failed agent response can still leave edited files. When switching to Codex is safe, what –fallback-model covers, how to resume.

Categories AI Coding Agents

Mac mini M6 vs M5 Pro for Developers: Is 32GB Enough?

November 12, 2026October 14, 2026 by jjeongil
Recreated Activity Monitor Memory tab on a Mac mini M6 with 32GB, with callouts on Memory Used, Swap Used and the memory pressure graph, process names and values hidden

With coding agents and Docker running, my 32GB Mac mini M6 reached the low 30s of GB. If your dev work fits in 32GB, get the M6; past that, the M5 Pro.

Categories AI Coding Agents

n8n Retry On Fail Sent It Twice: Why, and How to Fix It

November 2, 2026October 13, 2026 by jjeongil
How n8n Retry On Fail turns a timeout into a duplicate execution, compared with an idempotency key check that applies the request only once

n8n Retry On Fail re-runs the whole node, every item included, after a timeout, even when the first call landed. Why it duplicates, and how to fix it.

Categories Backend & Servers

Antigravity CLI Always Proceed Still Asks? 5-Run Test

November 10, 2026October 12, 2026 by jjeongil
Diagram showing Artifact Review as a separate flow, while a command passes through the Tool Permission mode and fine-grained rules, with an outside-the-sandbox boundary, to allow, ask, or deny

Use agy –dangerously-skip-permissions, but check your ask rules if it still asks; over 5 runs per mode, review mode stopped 4 actions I hadn’t approved.

Categories AI Coding Agents

RAG Retrieval vs Generation Failure: How I Tell Them Apart

November 18, 2026October 11, 2026 by jjeongil
Diagnostic flow from expected evidence to Top-k, final context, and answer, showing which RAG failure label applies at each stage

Wrong RAG answer? Find the right chunk in the top-k list, then in the prompt. In our 30-question run, heading paths halved retrieval misses; accuracy held.

Categories AI Coding Agents

AI Agent Calling the Wrong Tool? I Cut It from 6/20 to 1/20

November 1, 2026October 9, 2026 by jjeongil
Diagram splitting AI agent tool calling failures into four layers, tool selection, argument generation, call policy, and result interpretation, with the fix for each layer

My agent picked the wrong tool in 6 of 20 runs. Cutting tools didn’t help; overlapping descriptions were the cause. Adding ‘when not to use’ cut it to 1.

Categories AI Coding Agents

vLLM vs llama.cpp on a Single RTX 4090: Why I Went Back

October 31, 2026October 8, 2026 by jjeongil
Recreated Ollama API JSON response with prompt_eval_count and prompt_eval_duration marked as prefill, eval_count and eval_duration as decode, and the tok/s formula

On one RTX 4090 with one user, vLLM and llama.cpp generate at about the same rate; vLLM pulls ahead with many concurrent users. My trial and the numbers.

Categories AI Coding Agents
Older posts
Newer posts
← Previous Page1 … Page3 Page4 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