Episode 19 · Article on the episode

Prompt drift and the vanished click: an AI year in review

Which tool survived 2025, and what was merely the hot new thing of the week before last? A review along the three kinds of use that actually matter for most people.

By Mark Zimmermann · 15 Dec 2025 · 5 min read · Auf Deutsch lesen

A year in review without the politeness, sorted by what people really do: prompting and searching, generating images, and around Christmas generating music too.

The most uncomfortable finding concerns not a single tool but a property of all tools.

Prompt drift

The model carousel of Mistral, Grok, Claude and OpenAI overtakes itself on a weekly cycle. For everyday work the ranking matters less than a side effect that never appears in product announcements.

Established prompts that ran cleanly on GPT-4 suddenly produce unrequested extra information or entirely different blocks of text after the jump via GPT-5 to 5.1 and 5.2. The input is the same, the result is not.

Why prompt drift is unavoidable

A prompt is not an instruction to a machine but an input into a statistical system. It works because in this particular model it reliably triggers a certain reaction.

A new model is a different system. It was trained on different data, tuned differently and has a different system prompt sitting over it. That a prompt keeps working the same way is the coincidence, not the rule.

Three practical consequences follow from that. First: prompts on which something important rests need test cases with expected results, run through after every model change. Second: pin the model version in production workflows instead of switching automatically to the newest one. Third: phrase things so that the format is enforced, for instance through a prescribed output schema. An enforced format survives a version jump better than a finely crafted formulation.

This is exactly where skills have the advantage over prompts: they describe the task including the format, and they can be checked.

The click that fails to come

Both hosts see the biggest upheaval in classic search. Since Gemini has been sitting directly in the search box, the answers have been genuinely good for the first time since December 2025, after bumpy early months with questionable sources.

For website operators that is bad news: when the answer is delivered directly, the click through to the source fails to happen. In 2026 advertising is set to move into these results as well.

The practical consequence for everyone who publishes content is uncomfortable and unambiguous. If reach no longer comes from clicks, it has to come from something else: from content that gets quoted rather than summarised, from your own channels with direct access, or from offerings that an answer machine cannot replace.

At the same time Google is showing how broadly it can play its own advantage, with its own chips, its own cloud and its own devices: from experiments in Google Labs with personalised learning material, through NotebookLM, to the agreement under which Apple buys in a Gemini model in order to build a smarter Siri.

The voice assistants and their second chance

Apple and Amazon come off badly. Siri with ChatGPT access bolted on feels clunky because of latency and the same notice sentence every time. Alexa stays at ten to fifteen simple voice commands a day: weather, music, light, and thus a long way from a dialogue.

The hope is still justified. As soon as standards such as MCP and agent-to-agent protocols have matured and the devices at home have enough compute for local models, the established vendors are in a better position than any challenger. Their advantage is banal and hard to catch up with: their devices are already in the living room.

What shifted among the tools

A recurring annoyance of the year is interface experiments around reasoning displays: sometimes the thinking is shown, sometimes hidden, sometimes you get to choose between an instant answer and a research mode. Two answer variants side by side also strike both hosts as pressure to decide rather than as convenience.

Among the creative tools, by contrast, a lot moved overnight. Nano Banana turned Gamma for slides into an infographic and image machine with 4K output, and anyone who needs PowerPoint instead of Google Slides pushes the result through Manus.

An example from school shows the range: photographed notes go into NotebookLM, which builds infographics, explainer videos, quiz questions and flashcards for the next test out of them.

In programming, development ran from Lovable and Bubble via Replit to Manus, which generates applications with login, database and file upload from a few inputs.

Conclusion

The soberest statement in the review concerns not the tools but the precondition for using them. Anyone who wanted to get the most out of 2025 still needed insider knowledge: which subscription, which tool, which workflow currently fits.

That is the real barrier to entry, and it has risen rather than fallen. The tools have got better, the landscape more confusing.

Two things follow for your own work. Pin model versions in everything that has to run reliably, and put test cases alongside them. And check how much of your reach depends on clicks from search, because that source is drying up right now.

The large-scale delegation of whole goals instead of individual prompts remained a topic for the following year.