My old way was: whenever I had a question, I asked it, got a block of answer, and copied it into my notes. The result was that the answer looked useful, but nothing actually got done.
My current way is different: I hand a single"thing to do"to MuchSci 's conversation + agent, let it break it into steps and actually finish it, while I review at a few intermediate points.
The sentence I threw in
"Organize the three sets of electrochemical test data from this semester into a comparison table, mark the outliers, and add a paragraph of explanatory text that can go directly into a group meeting presentation."
There were no formatting requirements and no step-by-step instructions. What I was watching was how MuchSci 's conversation + agent would break it down next.
The five steps it broke out, and what I review at each step
Step 1: Confirm the input. It first asks back where the data is, how many sets there are, and what the units are. At this step I review whether the objects are correct——whether the three sets are the three sets I meant.
Step 2: Organize into a table. It consolidates values scattered across several files into one table. I review the column names and units.
Step 3: Find outliers. It provides the basis for deciding how much deviation counts as an outlier.
Step 4: Generate explanatory text. A readable explanatory passage that can also go directly into the presentation.
Step 5: Export files. One Excel, one Word document.
These five steps are exactly MuchSci the value its conversation + agent gives me: task decomposition. One sentence goes in, and what comes out is five executable actions, not a piece of general advice. I later found that the quality of task decomposition almost determines the quality of the whole thing——If it's broken down coarsely, every later step needs rework; if it's broken down finely, I only need to nod at the key points.
What's the difference between direct Q and A and conversation + agent
For the same sentence"help me look at this semester's data", direct Q and A gave me a block of analytical text: it was written quite decently, but I still had to make the table myself, prepare the figures myself, and write the presentation myself.
Switching to MuchSci conversation + agent, what I got was a table, a passage of explanatory text, and a file I could keep editing. The difference is not how smart the answer is; it's that it actually did the work.
My current usage is: for questions that need"thinking through", ask directly; for things that need"being made", hand them to MuchSci 's conversation + agent.
I write the request in three sentences
What to do, who it's for, and what counts as done.
If the request is written clearly, the task decomposition deviates less; if the last sentence is missing, what I get back will need rework. These three sentences are now my fixed format; whenMuchSci 's conversation + agent plans according to them, the probability of errors drops noticeably.
Three features I often combine
A complete task rarely uses only one feature. The combination I often use is:
- Finished documents: the ones that ultimately need to be submitted, such as PPTX, Word, use MuchSci office assistant agent to produce them; they are editable finished documents, and I can revise them once according to the template and send them.
- Keeping materials: conclusions found along the way and organized tables go into MuchSci personal knowledge base, so next time I write something I can pull them up directly instead of searching again.
- Literature support: where citation support is needed, use MuchSci paper search assistant to verify the DOI and citation information, rather than writing references from memory.
With the three features linked together, a task goes from"idea"to"submittable materials"without a break in between.
This is also how I now understand this kind of graduate-student AI assistant: it doesn't do the work for me; it breaks the work down to the point where I can review it at key points.