How generative AI is used in producing this site — what AI handles, and what it does not

How generative AI is used in producing this site

The question this page answers: Is this site's content written by generative AI? What role does AI play, where does human judgment come in, and can the content produced with AI assistance be trusted?

Background: the blog series and this section are separate things

This site has two distinct layers.

• Blog series Pt.0–Pt.25 — Blog posts published on microgroove.jp, researched and written by the author from September 2022 to January 2025 using primary sources (circuit diagrams, technical papers, academic journals, trade publications, patents, etc.), entirely by hand. No generative AI was used.
• This section — A set of pages newly structured and written based on the blog series and additional information obtained afterward, organized around readers' questions. It is built as a static site independent of the blog, and generative AI is used in the production process.

The blog series is a record of what a human being researched, thought through, and wrote by hand at the time. Some parts could be done more efficiently with today's technology, but I believe the research process itself has value, and I have chosen to preserve it as-is. This section is a reconstruction built on that foundation, with the help of generative AI.

The rest of this page describes specifically how AI is used in this section. Note that this reflects the workflow as of April 2026. Generative AI technology is evolving rapidly, and the tools and methods may change in the future.

Tools used

The primary tool is Claude Code (Anthropic). It handles everything from content structuring and drafting to technical implementation (Eleventy templates, CSS, build configuration) and English version creation — all within a single tool. The models used so far are:

• Claude Opus 4.6 (March – mid-April 2026)
• Claude Opus 4.7 (mid-April 2026 – present)

Within Claude Code, specialized sub-agents are used for different purposes. For example, an Explore agent is used for close reading of primary source PDFs and analysis of lengthy source material, while a Plan agent is used for designing major revisions or new pages.

Codex (OpenAI) is also used, with a clear division of roles between the two. Codex handles background research and analysis — close reading of primary source PDFs, organizing points of argument, and proposing candidate passages to incorporate. Claude Code handles drafting and implementation — producing draft text for FAQ and In a Nutshell pages, creating the English version, and implementing the site itself. Claude Code assigns subtasks to Codex, which carries them out and hands the results back to Claude Code; the author and Claude Code then use those findings to produce drafts and implement changes. All final wording and editorial judgment remain with the author.

What AI handles

Re-analysis of primary sources — The author read numerous primary sources (papers, technical documents, patents, etc.) while writing the blog series. In restructuring that material for this section, those primary source PDFs are fed primarily to Codex for re-analysis. Codex double-checks for anything the author may have overlooked or misunderstood, and proposes candidate passages and placements to incorporate. The author then reviews those findings and integrates them into the actual text (with Claude Code assisting in draft production).

Content structuring — The blog series was written following the chronological order of the research. When restructuring it into FAQ pages, AI collaborates with the author to reorganize content around readers' questions. Specifically, AI reads the relevant portions of the blog series and proposes structural options, which the author then selects from, adjusts, and refines.

Drafting — Once the structure is decided, AI generates a draft, which the author then revises — adding, modifying, and deleting as needed. No page has been published using an AI draft without revision.

English version creation — The Japanese version is completed first, and then Claude Code creates the English version. The author reviews and corrects as necessary.

Technical implementation — The static site built with Eleventy (templates, CSS, build configuration, deployment scripts) is implemented almost entirely by AI, with the author reviewing and approving the results.

What AI does not handle

Primary source research and selection — All decisions about which sources to rely on are made by the author. While AI can read and summarize materials the author specifies (blog post HTML, PDF papers, etc.), discovering, selecting, and evaluating the reliability of sources is the author's responsibility.

Final judgment on factual accuracy — Whether a historical statement is accurate is ultimately verified by the author against primary sources. AI can generate statements that diverge from the facts (so-called hallucination), and the verification step to prevent this is never skipped.

Handling of direct quotations — When quoting original English text, the primary source is always consulted directly. AI is never asked to generate quotations, nor are quotes reverse-translated from Japanese summaries back into English.

Distinguishing "fact" from "interpretation" — The judgment of what constitutes a "fact confirmed by historical sources" versus "the author's interpretation or conjecture" is made by the author. This distinction is fundamental to the site's credibility and cannot be delegated to AI.

Why use AI?

This is an individual research project. Restructuring and making bilingual the knowledge accumulated over two years of research required an enormous amount of work — more than one person could reasonably do alone.

By using AI, the author can focus on the essential judgments — what to write and what is accurate — while sharing the work of how to structure and how to implement with AI.

Specific quality control practices

• Project-specific canonical facts (often-confused points in EQ history) and writing rules (handling of direct quotations, terminology consistency, ban on reverse-translation, etc.) are maintained in dedicated reference files that the AI is required to consult before any work. This consultation step serves as the backbone of quality control, making it less likely that AI output will contradict facts or conventions already established in the project
• Every AI-generated draft is reviewed and revised by the author before publication
• When an AI-generated historical statement raises doubt, the author always returns to the primary source to verify
• Terminology consistency and factual accuracy have been corrected in AI output on numerous occasions
• The interaction with AI itself functions as a fact-checking process — AI proposes "may I write it this way?", and the author corrects with "that expression is inaccurate", in an iterative cycle

About this page itself

This page was also created following the process described above. AI produced an initial draft, and the author reviewed and revised the content.