by 逆瀬川ちゃん
16 min read
Hi there! This is Sakasegawa-chan (@gyakuse)!
Today I want to cover GPT-5.4, released earlier today by OpenAI: the instructions evolution readable from the Codex repo's source code, how to enable the 1M-context window, and benchmark comparisons with other models.
by 逆瀬川ちゃん
16 min read
Hi there! This is Sakasegawa-chan (@gyakuse)!
Today I want to take a deep dive into skill-creator, a skill officially released by Anthropic.
skill-creator is "a skill for making skills," and the structure of the skill itself is actually a treasure trove of best practices for skill design. On top of that, I'll compare it with agentic-bench (introduction article), a skill I previously built for automatic ML model benchmarking, and think about how to design "skills that orchestrate multiple processes."
by 逆瀬川ちゃん
14 min read
Hi there! This is Sakasegawa-chan (@gyakuse)!
Today I want to take on the "unsolved problem" I left dangling at the end of the previous post: the attention competition problem between skills. When you end up with 10 or 20 skills, are they really firing correctly? Did improving one skill's description silently hurt another one? I built a skill that audits this at the portfolio level automatically: skill-auditor (currently Claude Code only).
by 逆瀬川ちゃん
20 min read
Hi there! This is Sakasegawa-chan (@gyakuse)!
Today I want to share the story of a speech recognition model I built from scratch that outputs only hiragana. Why hiragana instead of mixed kanji-and-kana text, why Whisper didn't work for me, and how combining it with an LLM produces a practical voice dialogue system: I'll walk through the background, the implementation, and the evaluation results in one go.
by 逆瀬川ちゃん
7 min read
Hi there! This is Sakasegawa-chan (@gyakuse)!
Today I want to summarize, using the actual production of spark-banana as an example, what I learned about making product promo videos with Remotion and how I packaged that knowledge into a Skill.