Fatigue is real and the industry mostly pretends otherwise
Faster individual tasks lead to more tasks, not fewer
Your manager's expectations and your own will both adjust upward to fill the new capacity
Using an LLM shifts you from creator to reviewer
Evaluative work (judging output, spotting errors, deciding what to trust)
drains you faster than generative work does
If you outsource your first-draft thinking long enough,
your ability to reason from scratch degrades,
much like navigating without GPS after years of relying on it
So spend some time each day working through problems without the LLM
The prompt spiral is real
If you cannot get something 70% usable in two or three attempts,
write it yourself
The goal is to finish the analysis, not to get the LLM to produce perfect output
Treat every LLM response as a draft the moment it appears
Perfectionism plus probabilistic output is an exhausting combination
"Good enough to work from" is a legitimate bar
Keeping a log for two weeks---task, used LLM (yes/no), time spent, confidence in result---will
tell you more than any general advice
Where to go next
We wish we knew…
Exercises
Name one task from today where an LLM saved significant time
Estimate how long it would have taken without it
Name one task from today where you had to correct the LLM
Describe what you needed to know in order to spot the error
Find an energy cost estimate for a single LLM API call
Compare it to the cost of sending an email or running a SQL query
Identify one part of your analysis where you are not confident in the result
Write down exactly what you would do to check it
Find the official documentation page for the tool in the stack you feel least confident about
Identify one concept you want to learn next
Write a two-sentence policy for your own use of LLMs in research