The post is a month-by-month record of one practitioner’s 2025 output, written because the author had occasion to compile it for work. What makes it more than a list is its opening premise — that 2025 will be remembered as the year the way software is made changed substantially — and the fact that the author’s subject for the preceding two decades has been automated testing and test-driven development, so the year’s changes landed directly on his material. He also states plainly that the post is entirely his own writing, with AI involved in neither its prose nor its structure.
Read as a timeline, it is unusually specific about when things turned. He describes his own mood at the end of 2024 as gloomy, watching coding agents grow capable and the territory where humans write code start to be cut away, and wondering whether people would write code at all. February is where he puts the decisive month, naming three publications from it — two on consecutive days and a third later that month — and summarising their combined effect as the human moving from the driver’s seat to the passenger seat and the AI moving the other way. In March, on hearing that a coding agent usable from the CLI had appeared, he starts using Claude Code straight away; with his editor no longer pinned by the agent he returned to Emacs, and switched terminal too. In May the pricing change lands, and he identifies it immediately as decisive: with a flat monthly rate rather than per-token billing, he could iterate without watching the meter, and what he could see changed.
The professional consequences run alongside. New-hire training had to be restructured, because he had to decide how to teach Red/Green TDD in an era when agents write the code; his solution was a two-part format teaching the current situation and then the enduring design fundamentals behind it. He notes that in demonstrations, new engineers responded more strongly to watching a specification be argued out with a model than to watching code be written test-first. The year’s most-read deck of his was a completely new talk on software development in the age of AI, written and rewritten in a green room the morning it was delivered. In December he ran a Spec-driven development training course for the first time, and reports the difficulty that makes such courses hard to guarantee: an agent’s behaviour is probabilistic, so the quality of the experience is hard to assure.
Key Points
- The post’s framing claim is that 2025 will be remembered as the year the way software is made changed substantially, and the author states it was a particularly memorable year for him personally.
- The author states explicitly that the entry is 100% his own writing and that AI was involved in neither its content nor its structure.
- His recorded state of mind at the end of 2024 was unsettled: coding agents were growing rapidly capable, the area in which humans write code was beginning to be pared away, and he records wondering whether humans would write code at all, while noting he loves coding.
- He dates the decisive shift to February 2025, naming three publications from that month — including one he credits with making Vibe Coding a talking point — and summarises their combined effect as the human moving from the driver’s seat to the passenger seat while the AI moved from passenger to driver.
- He began using Claude Code in March 2025, reporting that he started as soon as he heard a coding agent usable from the CLI had appeared; the consequence he records is that his editor was no longer constrained to VS Code, so he moved back to Emacs and kept that configuration for the rest of the year.
- He identifies the May 2025 change that made Claude Code available under a flat-rate subscription as decisive, on the stated grounds that it changed the incentive structure: without per-request billing to watch, he could experiment aggressively, and what he could see changed as a result.
- Before Claude Code gained a planning mode, his workflow was to settle the specification in a chat interface first and only then move to the agent — and he reports new engineers in his training reacting more strongly to that specification discussion than to writing code test-first.
- Restructuring his standing new-hire training was a genuine difficulty for him: he records not knowing how TDD should be taught, or what career to describe, in an era when coding agents write the code, and settled on a two-part format covering the present situation and then the underlying design fundamentals.
- He reports a client explicitly commissioning half-formed material — asking him to speak on TDD in the age of AI even if it was not fully worked out — and treats that as a challenging but welcome kind of request.
- A practitioner’s finding is recorded from June: that instructing Claude Code to follow this author’s own recommended approach is markedly effective at getting it to do TDD. The author’s reading of why is that the terms TDD and test-driven development spread so widely that their meaning thinned, becoming confused with automated testing and test-first, which affected LLM training data — and that supplying a person’s name gives the model a concrete reference point that narrows it to a specific style (see AI に t-wada の TDD を頑張らせる).
- His own open-source work that month went the other way: he developed a new output format for his assertion library aimed at being LLM-friendly, and states that no AI wrote that code — he wrote 100% of it himself. His stated motivation is that an era in which agents run tests frequently and judge the results needs both ease of writing tests and informative failure output.
- He ran his first spec-driven development training course in December 2025, and reports the problem it exposed: because an agent’s behaviour is probabilistic, assuring the quality of a training experience built on one is difficult, and he expects to keep searching for a workable format.
- His summed 2025 output is 53 talks, 12 training courses, 10 panel discussions, 6 podcast recordings, 5 question-and-answer events, 4 interviews and one book published as supervising translator, with his talk on software development in the age of AI given 15 times.
- Two of his 2025 decks were nominated in the year’s most-viewed presentations on the platform he publishes them to: the one on software development in the age of AI and one on judging technology choices.
- His closing claim is the one with the widest reach: that after nearly twenty years of advocating automated testing, test-first and TDD in Japan, 2025 was the year automated testing became most widespread, because AI agents substantially reduced the two obstacles in its way — learning cost and implementation cost — and that automated testing has finally become ordinary.
- He immediately qualifies it: the accompanying Semantic Diffusion is severe, and his stated remedy is simply to keep explaining, patiently, again.
- The post closes on a reversal of its own opening: the gloom of January 2025 had lifted, replaced by a readiness to play under the new rules of the game.
Context
This is a personal retrospective rather than an argument, and its evidential status varies by claim: the counts of talks and courses are the author’s own records, while the assessments — which month was decisive, what the pricing change meant, whether automated testing has become ordinary — are his judgment from one vantage, that of a Japanese testing specialist whose consulting and training work put him in front of many organisations during the year. That vantage is the post’s main value and also its limit: the claim that automated testing became widespread in 2025 is offered as what he feels rather than as measurement, and he says so. Several of the publications and events he credits with turning the year are linked rather than described, and their own contents are not established here. The term-thinning he names at the end is treated as its own subject under Semantic Diffusion, and the practice whose adoption he tracks under Agentic Coding.