People who read the classic book "Essentialism" are good at discarding ideas returned by AI. They generate 20 options, read them, and keep only 3. As a way to use the book, this is certainly correct.
However, I think people who have become busier after introducing AI are probably still practicing this method too faithfully.
The more you type "Give me something good" or "Make it better" into ChatGPT or Claude, the more your "discarding work" increases, sucking up your time.
"Essentialism" was published in 2014 and has sold over 600,000 copies. It includes the "90% Rule": decide on one most important criterion, score options out of 100, and discard anything below 90 as if it were zero.
In other words, the book uses this tool when choosing from lined-up options. But with AI, you can set the criteria for discarding (below 90%) *before* generating.
The 90% judgment is excellent, but I believe shifting the timing "forward" makes it even more effective.
Until recently, I also asked AI to "provide 30 diverse options." Then I spent time reading and choosing, almost thinking that AI only returns obvious answers... But upon reflection, I simply hadn't written down the "criteria for discarding" beforehand. So now, I firmly place 1-2 sentences of discarding criteria at the very beginning of my prompt.
Even if the conditions for "something good" aren't yet articulated in your mind, haven't you roughly decided the conditions for "things you don't need to see" before asking?
You might think, "I don't have those criteria..." In that case, starting from "Give me candidate criteria for things to exclude" improves the rationality of AI usage.
When instructing AI, write what you don't want to see before writing what you do want. The sentence pattern is just two sentences: "Don't include [X]. On top of that, [Y]."
Here are examples categorized by three purposes for asking AI:
<Generating Ideas>
"Give me something good"
→ "Don't include anything that exceeds this quarter's budget. On top of that, give me 20 options."
"Give me 10 catchphrases"
→ "Don't include ones that just rephrase the product name. On top of that, give me 10."
"Please give me project ideas"
→ "Don't include anything we can't manage with our current headcount. On top of that, list them in order of what we can try next week."
"Give me title options"
→ "Don't include titles that promise content not in the body text. On top of that, give me 5 shortest ones."
"Give me something good" seems to be interpreted by AI as "give me everything you can think of." If you use this after reading the 90% rule, you end up with the job of reading 20 options.
<Researching>
"Look into [Topic]"
→ "Don't include information found on the first page of the official website. On top of that, show me only differences from us."
"Check competitors"
→ "Don't include companies larger than ours. On top of that, show me 3 companies in the same price range."
"Tell me more details"
→ "Don't repeat the answer from the first time. On top of that, provide one opposing viewpoint."
"Prepare for a meeting about [Issue]"
→ "Don't include agenda items we won't decide today. On top of that, show me one thing to decide this week and its materials."
In research, the more known information is mixed in, the longer the reading time becomes. By excluding the known first, you only read the delta.
<Editing>
"Make it better"
→ "Don't make edits that change the meaning. On top of that, adjust only the phrasing."
"Make it easier to understand"
→ "Don't add explanations of terms. On top of that, split long sentences into shorter ones."
"Check this"
→ "Don't include points that only trigger from one perspective. On top of that, show only issues where two or more perspectives point to the same spot."
"Fix it again"
→ "Don't include sections fixed last time. On top of that, leave a column showing the fixed parts and reasons."
If you just say "fix it," AI interprets it as permission to rewrite the entire text. Worse, you end up with extra work comparing the before and after versions.
This happens because while you have "criteria for discarding from choices," you haven't prepared "criteria for not creating candidates in the first place."
Although the content of these rewrites differs, the action taken at the beginning is the same.
Decide "what to discard so it doesn't even enter the choice pool" before handing it to AI,
Then add the conditions for "good stuff" afterwards.
How many lines of "results that didn't need to be output" were in your usual AI responses?
Fast workers leverage AI starting from this "point of narrowing down candidates."
I also explain timely news and hints for tomorrow's work structurally and clearly.
https://x.com/smark_x/status/2099383008537297309
https://x.com/smark_x/status/2096168360103088156
• Overthinking prevents you from taking action
• Perfectionism or inefficiency leaves you short on time
• When hitting a wall, you want to blame it on "lack of sense"
• You introduced AI, but work hasn't actually gotten easier
I had all these issues myself, but by systematizing with "logical thinking," my work became easier. I post daily, so please follow if interested. → @smark_x
