Welcome to Prompt, Tinker, Innovate—my AI playground. Each edition gives you a hands-on experiment that shows how AI can sharpen your thinking, streamline your process, and power up your creative work.
This week's playground: The tool you pick before you've typed a single word
Last week's newsletter was about the three dials that impact usage limits—model, effort, speed. Turn any of them up, and you burn through your usage limit faster. That's still true. But there's another variable that comes before all three: which tool you choose in the first place.
Regular chat, Research tools, and agentic tools like Claude Cowork, Microsoft Copilot Cowork, ChatGPT Work, or Gemini Spark are built for different kinds of jobs. They can require very different amounts of context, reasoning, tool use, and compute, and those differences can directly affect how quickly you use your allowance.
The simplest way to think about it: do you need an answer, an investigation, or action?
Regular chat
Major AI platforms have a version of this.
This is your baseline. You ask, it answers, sometimes it checks the live web first. It's reactive—you're driving each step. What each message costs is still shaped by the model dials: which model you picked, how hard you told it to think, and how fast you wanted it back.
Other things can matter too—conversation length, files, web searches, and tool use—but chat typically involves less autonomous work than Research or agentic modes. It responds to the turn you gave it rather than taking a whole task and deciding what to do next.
Good for: quick lookups, drafting, back-and-forth thinking, anything that fits in a normal conversation.
Research / Deep Research
Major AI platforms have a version of this.
This tool goes off and investigates on its own: searching the live web across dozens of sources, reading pages and PDFs, cross-referencing what it finds, and returning a structured, cited report. It can run for minutes, sometimes tens of minutes, because it's doing real multi-step research rather than producing a single conversational response.
Here's what that actually costs you: Research can search and process dozens or even hundreds of sources, pulling relevant material from pages, PDFs, images, and other sources into the work before it writes back. You're not just paying for the answer—you're paying for the investigation that produced it. Different providers meter that work differently, but more research generally means more of your allowance consumed.
Ask it something chat could've answered in one message, and you've spent all that research effort on a question that didn't need investigating. If it's a basic question, don't reach for Research. Save it for when the answer genuinely requires investigating across multiple sources.
Good for: competitive research, literature reviews, anything where "compile everything relevant on this" is the actual ask.
Agentic work
Claude Cowork, Microsoft Copilot Cowork, ChatGPT Work, Gemini Spark.
This is where the cost difference can get large. In an agentic tool, your one message can turn into many autonomous steps: planning the work, reading files, calling tools, checking results, revising outputs, and deciding what to do next. Each step can add context and computation. Regular chat can also use tools or perform multiple internal steps, but agentic mode is built to keep going without you directing every move.
That scaffolding—system instructions, tool access, file context, and intermediate work—can add overhead before the task itself gets complicated. So even a basic question in Cowork or Work can consume more of your allowance than the same question in chat simply because you've chosen a mode designed to do more.
Where that overhead really compounds is complex, multi-step work. The longer the task runs, the more context it needs, and the more tools it calls, the more usage can climb. OpenAI makes the same point about Work: usage varies based on things like task complexity, context, reasoning, speed, and tools.
The point isn't that agentic tools are wasteful. It's that they can be expensive when you use them for work that doesn't need an agent.
Good for: work that actually spans multiple steps, files, or tools—building a folder structure, filling out a spreadsheet with working formulas, managing a workflow across apps, or carrying out a task end to end without you doing each step by hand.
Your AI experiment: Match the task to the tool, before you open anything
👉 Time to tinker: Before your next task, stop and answer one question first: does this need an answer, an investigation, or action? Then open the tool that matches—chat for the answer, Research for the investigation, agentic for the action.
💡 Pro tip: Like last week, check your usage meter where your plan exposes it before and after big tasks. That's where you can start to see the difference between a Research or agentic task and a standard chat conversation.
What did you discover?
Were you opening a higher-usage mode out of habit for something chat could've handled? Or did switching modes save you real usage you didn't know you were spending? Tell me what surprised you.
Until next time—keep tinkering, keep prompting, keep innovating.
📩 Not subscribed yet? Hit the button at the top. You won't want to miss what's next.
