engineering

"I upgraded my Claude Token Counter tool to add the ability to run the same count against different models in order to compare them."

"By shifting report generation to the Python side, we were able to suppress credit consumption and execution time."

"Vol.4 ではゼロから再設計し、6つの AI エージェントが役割分担しながら自律的に動くマルチエージェントオーケストレーションに生まれ変わった。"

"LLM itself does not have memory. It really 'shouldn't remember' what was talked about in the immediately preceding conversation. Despite this, when chatting with chat apps like ChatGPT or recent agents, it feels as if the other party remembers you and possesses memory."

"The strongest model is updated every 1-2 weeks, and sticking to a specific model no longer makes sense. Model selection is no longer a differentiator—what is being asked instead is whether you have the design ability to master Generative AI."

"The problem is that every time a review is performed, the Agent starts from a completely blank slate. A real-life reviewer has context, such as 'I made the same comment on this file last week' or 'This project prioritizes internal coding standards over PEP8.' This time, to solve that problem, we implement a two-layer structure: short-term memory (conversation history) and long-term memory (JSON file)."

"Love how Gemini personalizes the response by mixing up coding terms with snacks!"

"AIエージェントのおかげで圧倒的コーディング力を手にいれ、どんどんコードを量産する。"

"私が出した結論はやっぱり、自分で書くべきところ、ないしはアウトプットするところは残しておいた方が良いというのが暫定の結論である。出来るだけ最小単位で AI を使って、コアの部分は自分が実装してレビューしてもらうのが良いのではないかと考えた。"

"Before and after. Done in 5 mins. One-Shotted. I'm genuinely stunlocked"

"schema markup alone shifts precise information extraction from 16% to 54%. That's not a marginal gain — that's the difference between being cited and being invisible."

"Structured context mattered more than model size in many cases."

"By introducing the right Skills, the quality of the UI generated by Claude Code changes surprisingly. Skills transform it from just a 'working UI' to an 'intentionally designed UI.'"

"In the AI-native era, if the preparation is sloppy, the AI will sprint at high speed in that sloppy state and reach a point of no return. This creates an asymmetry where poor preparation forces you to pay 2-3 times the effort in post-processing."

"I gave it a 40-line order processing function, and what came back was a 120-line code with the Decimal class, dataclass, logging module, type hints, and Strategy pattern added."

"「初回ヒアリングだけでも自動化できないか？」という発想から、このシステムが生まれました。"

"no matter how i phrase it in the instructions, how many times i repeat the rule not to use quotes, and which LLM i use, i have failed to prevent any of them from using the so-called scare-quotes."

"As the conversation grows longer and the history expands, the model's attention becomes scattered, leading to decreased compliance with constraints, increased costs, higher latency, and hitting context limits."

"By describing things as 'Absolute Rules' in CLAUDE.md like this, Claude Code will spontaneously start checking for them. For example: 'Never commit files containing API keys, passwords, or tokens (.env, credentials.json, etc.).'"

"If you want Claude Code to feel like a senior engineer living inside your repo, your project needs structure."

"Prompts should be verified by having the AI that uses them run a correction loop. This creates the same structure as TDD, where the AI using the prompt is the only one capable of providing the criteria for prompt evaluation."

"Harness: If you block it with a linter, it physically cannot be used."

"The context window is huge, but as it swells, the AI's attention becomes scattered. It's not that a larger context makes it smarter; if it gets too long, performance degrades. AI is truly looking at the entire conversation history every single time."

"AI agents, especially the type that autonomously rewrite code (like Claude Code or Gemini / Codex, etc.), will always face 'losing context' and 'wasting tokens'."

"Now it feels like training is just one piece, and everything around it is where most of the difficulty is."

"An agent harness is an 'execution platform that wraps around the outside of the Large Language Model (LLM) agent.' A 2026 survey paper reported five cases where performance improved up to 10x solely through harness configuration without changing the model."

"Claude Code Game Studios is an Open Source project that allows you to run a full-fledged game development studio entirely on your own using 49 specialized AI Agents and 72 workflow skills, essentially functioning as an 'AI studio simulator' with the exact same organizational structure as a real game studio."

"Anthropic isn’t vibing with me today 😢"

"Claude Code is not just used as 'a single AI assistant', but can be used as all members of a team."

"Markdown is a format markup language... In recent years, Markdown documents have often been used for model training, thereby becoming the native language of AI models."