A VS Code extension with 4 installs takes on a real problem: vague prompts that waste AI tokens

By Zendoric · October 4, 2026.

iNoU Agent is a Visual Studio Code extension published by independent developer Mauricio Gracia Gutierrez. According to its own listing on the Visual Studio Marketplace, its core idea is twofold: on one hand, to score locally the clarity of every instruction the user writes (what it calls the "Specifi-Meter") and, if the prompt is too vague, to pose a multiple-choice question instead of sending it straight to a language model; on the other, to automatically route each task to the most suitable model —a fast one for simple questions, a more powerful one for refactoring or multi-file debugging— combining the user's own Google Gemini and Anthropic Claude keys or integration with GitHub Copilot. Code changes always arrive as a diff that the user reviews and approves, with a read-only mode that blocks any edit until write mode is explicitly enabled.

It is a "BYOK" (bring your own key) tool: it does not include model access of its own, but depends on the user connecting their own paid credentials. According to that same listing, as of this publication the extension had accumulated 4 installs and zero ratings; the Hacker News thread where it was shared had 1 point and 1 comment. These are the figures of a project that has barely left the garage, not of a product with traction.

As industry context, the agentic coding assistant niche —tools that not only autocomplete but plan, touch several files and run tasks on their own inside the IDE— has filled up with competitors over the past two years, from the major providers' built-in assistants to a long list of independent extensions and forks. In that market, an interesting, well-described technical proposition like iNoU's is not enough on its own to stand out: distribution, prior trust in the author and the fit with an already installed workflow weigh as much as or more than the quality of the design.

Our reading is that the relevant takeaway from this piece is not the product itself —too small to draw market conclusions from— but the problem it tries to solve, which is real and documented by those who use AI assistants every day: an ambiguous prompt not only produces a mediocre answer, it also spends tokens (and money) in the wrong direction, which then has to be undone and redone. That a niche tool designs an explicit mechanism to stop that waste before it happens, instead of merely generating code faster, points to a maturing in the design of these tools: the question is no longer just "how smart is the model?", but "how do I keep the model from working on a badly framed request?".

Also revealing is what the extension treats as non-negotiable: that the user review every diff before applying it, and that write mode be enabled on purpose. It fits an idea we repeat at Zendoric about the future of technical work: AI does not replace the judgment of whoever approves a change, it makes it more necessary. The more tasks a programmer delegates to automatic agents, the more valuable the ability to review, question and decide what goes into the repository becomes, rather than typing out the solution line by line. A four-install project does not change the industry, but it does illustrate where the design of the tools that will change it is heading.