Many AI features, like Gmail's AI assistant, feel useless because they don't allow users to edit system prompts, constraining the AI models they're built with
Millions Of Email Users Now At Risk Of Attack Mastodon: Dare Obasanjo / @carnage4life@mas.to : This blog captures my frustration with AI tools for work. Microsoft and Google are pushing tools to help you write email or documents but what I need is a help reading and combining information across documents, email, source code and chat. — That would be really useful. Not CoPilot for email. … @raganwald@social.bau-ha.us : A very interesting essay about AI being in its “Horseless Carriage” era. — https://koomen.dev/... This rings true: In the early days of graphical computing, “Desktop Publishing” was a massive, disruptive market. — But in the end, using computers to publish pieces of paper was horseless carriages all over again. … X: @garrytan : AI builders: Let the people write system prompts. https://koomen.dev/... [image] Aaron Epstein / @aaron_epstein : Super smart insights from @koomen: “In the new world I don't need a middleman to tell a computer what to do anymore. I just need to be able to write my own System Prompt, and writing System Prompts is easy!” Pete Koomen / @koomen : https://koomen.dev/... Gmail's AI assistant wraps google's incredible Gemini model behind a UI that forces me to explain how to write an email every time I want one written. Ben Thompson / @benthompson : This is excellent Mike Knoop / @mikeknoop : We found this idea of user-owned system prompt is a requirement, not an option, for Zapier Agents. There is no “god prompt” that works reliably across all use cases. The only way to get Agents reliable enough to deploy is let end-users locally steer to their specific use case. Pete Koomen / @koomen : I wrote an essay about Gmail's useless email-writing AI assistant: “AI Horseless Carriages”. link and TLDR in thread [image] Pete Koomen / @koomen : The future is a world where I don't have to spend time doing mundane work because agents do it for me. Where I'll focus only on things I think are important because agents handle everything else. Where I am more productive in the work I love doing because agents help me do it. LinkedIn: Chris Kong : One of the best things I've read lately on how we misuse AI: — We're still building “horseless carriages” instead of rethinking the road. … Bikram Dahal : For a while, I've been skeptical about the integration of AI in various products. Even in my own experience trying to implement AI features, few feature felt forced or artificial. … Daniel Fenton : I highly recommend this article to anyone thinking about how to reinvent products in the era of AI. Too much of what you see today feels like AI chat is haphazardly bolted on. … Sriram Natarajan : Most product folks (including Google PMs) are still learning on how to build AI-driven products that truly delight users. … Stephan Spijkers : “Most AI apps should be agent builders, not agents.” — I was in a call today being frustrated (sorry Tomas Hesseling) … Forums: Msmash / Slashdot : YC Partner Argues Most AI Apps Are Currently ‘Horseless Carriages’
Context & Ripple Effects
This criticism extends earlier doubts about whether generative AI features deliver durable value rather than novelty, including warnings that the technology could disappoint if products are built around overly sweeping assumptions about generative AI’s uncertain practical payoff. It also echoes reports of AI-generated workplace output that felt hollow rather than genuinely insightful in early style-generation tools.
The dispute is less about Gemini’s underlying model than about product control: Google’s Gmail interface keeps the governing instructions fixed, while the account here argues useful work assistance needs persistent, user-defined behavior across information sources.
First-order effects
- Gmail users remain confined to Google’s preset assistant behavior and must restate preferences in individual interactions instead of maintaining an editable instruction layer.
- Google’s email AI is judged against a broader task—synthesizing material across email, documents, code, and chat—rather than merely drafting messages.
Second-order effects
- Tools that expose user-owned system prompts, such as Zapier’s agent approach described here, gain a clearer differentiation point for teams that need repeatable, specialized workflows.
- Microsoft and Google face pressure to make workplace AI more configurable if fixed, writing-oriented assistants are seen as less useful than tools that operate across a user’s work context.
Third-order effects
- If user control becomes a central measure of agent usefulness, workplace AI may shift from isolated drafting features toward configurable, workflow-native work surfaces.
- The split between polished default assistants and user-steerable agents could determine which vendors become systems of record for AI-assisted work; that outcome depends on whether added control improves reliability without making products harder to use.
The trend: Enterprise AI is moving from generic embedded copilots toward configurable agents that retain user intent across workflows.