Pick Your AI by the Task, Not the Hype (A Simple Routing Framework)

When choosing an artificial intelligence tool, focus on the task at hand rather than the hype surrounding the model. The right question to ask is not which model is the smartest, but what specific failure modes will the task encounter? For instance, a marketing email requires a model to avoid a robotic tone, while a migration script needs to avoid a failed build, and a finance sheet needs to avoid an off-by-one column error. Each of these tasks has distinct failure modes, and therefore, distinct tools.

Before opening a chat window, classify the task into two components: the type of work and the potential failure mode. For example, long-form writing, editing, and tone require a general chat model with long context, such as Claude, GPT, or Gemini. Code, refactoring, and debugging benefit from a coding-focused model wired into your editor or terminal.

Data, spreadsheets, and analysis are best served by a model that can write and execute code, like Python execution. Images, mockups, and design require a dedicated image model, while research and current facts are best handled by a tool with real-time retrieval and citations.

Remember that tool names are not set in stone; rankings shift every few months. Treat them as families of tools rather than gospel. The core of the decision-making process lies in matching the tool to the failure the task cannot tolerate. If the task has a hard check, like compilation or accurate numbers, pick the tool that connects to that check. For example, a code-executing tool that can run tests on a repository will outperform a smarter model that guesses at files it cannot read.

The second principle is to use the general chat models as generalists for open-ended work with soft criteria, like writing and brainstorming. However, avoid using them when the task has mechanical ground truth, as they might produce confidently wrong results. A quick example is providing a plain chat model with the task of summing column C where the region is EU over pasted data.

While it will often produce a confident, slightly wrong number, a code-executing tool will write the exact code needed to perform the task correctly.

To validate your choice, run a small bake-off on tasks that you actually own. This costs an afternoon but ensures that the result reflects your work instead of an average of everyone else's. Score on what you care about: correctness, edits needed, and time saved. Keep a note of which tool won each category. This note is your real leaderboard, and you should re-run it every few months or when a tool you use ships a major update.

While categories may blur, and your real work may be mixed, running this framework is cheap and helps you stay on top of the best tools for your specific tasks.