Save your prompts, decode a text, make a hard call

The Take

The best prompts this week make the model push back before it helps. They don’t answer your question right away. They make you say what you actually want first & that bit of friction is the whole point.

The easy way is the opposite. You tell the model what you want in a word or 2 & it works.

That’s faster.

But it’s also the weaker move. It only works when you already know exactly what you want.

Most of the time, you don’t. That’s why you opened the chat in the first place.

The prompts worth saving this week don’t take you at face value. They make you sharpen the ask before the model answers.

Seen This Week

1. Save your best prompts as fill-in-the-blank templates

Most people collect great prompts & never use them. They forget, or can’t find them again. They’re buried in a note somewhere, so rewriting a worse version from memory beats digging it up or trying to remember where you saved it. The fix is to stop saving prompts as finished text & save them as templates with blanks you fill in.

Here’s the one I reach for most, a blunt critique of anything before I send it.

Act as a demanding {{role, e.g. senior editor / hiring manager}} reviewing my {{thing}}.

What it's for: {{goal}}.
Who it's for: {{audience}}.

Give me:
- The 3 weakest things, worst first, with a specific fix for each.
- One thing that's working, so I don't break it.
- A score out of 10 and the single change that would raise it most.

Be blunt. I want it better, not validated.

WORK: {{paste it}}

Why it works: You write the prompt once & reuse it, filling in the blanks instead of retyping the whole thing from memory. OP wraps the blanks in curly braces, but it doesn’t matter what you wrap them with, the model gets it. Mark the parts that change, fill those in each time, leave the rest alone.

One gap: A template is only as good as what you put in the blanks. Fill {{thing}} with “my email” & the feedback is so generic it fits any email.

Simple fix: Get specific. Fill it with “the opening 2 lines of my cold email to a VP” & the notes get sharp. “My email” gets a shrug.

Want to go deeper: The point isn’t this one prompt, it’s the habit. Here’s a second template so you can see the pattern, a rewrite that keeps your voice.

Rewrite this {{content type, e.g. email / bio / paragraph}} to be more {{quality, e.g. concise / warm / direct}}.
Keep my meaning and my voice. Don't invent facts.
Give me 2 versions, one safe and one bolder, with a one-line note on what changed.

TEXT: {{paste it}}

Every time a prompt works well for you, stop & turn the changeable parts into blanks before you move on. A few weeks of that & you have a library you fill in instead of rewrite.

Original post by u/Ok_Negotiation_2587. Link

2. Decode a tense message before you reply

Everybody has gotten that message that makes their chest tighten. The reflex is to fire back fast, which is exactly how it escalates. Before you reply, this prompt reads the message for the real problem hiding under the tone, then hands you a calm first line.

A message I got feels confrontational, and I want to understand it before I reply.

Their message: [paste it]
My relationship to them: [who they are]

Tell me:
1. The real fear or need driving their tone.
2. The actual problem they want solved.
3. One calm opening line I can use to lower the tension right away.

Why it works: A hostile message has 2 layers, the tone & the actual ask. Most people react to the tone & miss the ask. This pulls them apart, then gives you a first sentence that drops the temperature so you can deal with the real problem instead of the heat around it.

One gap: It trusts the message at face value. If the person is being vague on purpose, the need the model hands you is just a guess. It only sounds certain.

Simple fix: Add the backstory the model can’t see, what happened before this message landed, so it reads the situation & not just the words.

Want to go deeper: Once it decodes the message, paste back: “Now write me 2 replies. One that answers the real need head-on, one that buys me a day without sounding cold.” You get the read & the response in the same sitting.

Original post by u/EQ4C. Link

3. Make a hard decision without your bias running it

Most hard decisions are already made before you finish thinking them through. You lean one way, then you look for reasons to back up the side you already picked. That’s confirmation bias. This chain breaks that loop by making the model argue with you before it helps you decide.

Step 1, Decision Mapper: You are a neutral decision strategist. I have a hard decision. Do not give me a recommendation yet.
My decision: [describe what you're deciding].
What I'm leaning toward: [be honest].
Why I think I'm leaning that way: [your reasoning].
Do the following:
No advice yet. Just clarity.

Step 2, Devil's Advocate: Now argue against the option I'm leaning toward, not to change my mind, to stress test it.
- The 3 strongest reasons my pick could be wrong.
- The realistic worst case if I go that route.
- What a smart person who disagrees would say.

Step 3, Decision Framer: Using everything above, help me make the final call.
- The real trade-off in one sentence.
- What this decision says about what I actually value.
- A recommendation, with the one condition under which you'd change it.

Why it works: This prompt splits the job across 3 turns. Ask one prompt to help you & argue with you at the same time & you’ll get slop. Step 2 argues against the side you favor. Step 3 forces a clear answer instead of a comfortable one. You hear the case against your choice before anyone helps you make it.

One gap: Step 1 is supposed to set that all up, but if you look at it, it says “Do the following:” & then lists nothing.

Simple fix: Replace that empty line with real instructions. This version tells the model what to produce & the brackets make it a template you reuse.

Step 1: I have a hard decision: [decision]. I'm leaning toward [option] because [reason]. Don't advise yet. List the assumptions behind my lean, the facts I'd need to check, and the part of this I'm avoiding.

Want to go deeper: After Step 3, ask one more thing: “What single fact, if I checked it right now, would most change your recommendation?” That’s the thing to go find before you commit.

Original post by u/LouiszzZ_. Link

The Teardown

Slash commands are doing less than you think

Here’s one of the newest prompting techniques making the rounds on Reddit this week. Someone posted a library of 60 slash commands, /eli5, /redflags, /steelman, things you type into the chat box to trigger a behavior, like a secret menu hidden inside the model. OP sorted all 60 by job into a list that’s genuinely useful. The question is whether the slash is doing any of the work.

/steelman - Present the strongest possible version of an opposing viewpoint.
/eli5 - Explain something in simple terms as if to a 5-year-old.
/ownership - Clarify what is truly your responsibility versus someone else's.
/redflags - Identify hidden risks, weak assumptions, or likely failure points.
/boothmath - Evaluate a booth or thrift item by profit, labor, storage, and style fit.

OP’s claim: these aren’t official, but the model recognizes them & runs the action.

Prompt Teardown

The slash isn’t the magic. The word is.

Type /steelman & the model already knows what steelman means, give the strongest version of the other side. Drop the slash, type “steelman this,” & you get the same answer. It was reading the word, not running a command.

But /ownership? Ownership of what? The model has no idea what you want from that. It does something only because OP wrote a line next to it, “Clarify what is truly your responsibility versus someone else’s.” That line is the prompt.

Go further down the list & it gets obvious. /boothmath means nothing to anyone but the person who coined it. There’s no word in there for the model to read. It works only because the description spells out profit, labor, storage & style fit. Take the description away & the slash is just noise.

OP wrote these for ChatGPT. Try them in Claude & the slash can even backfire. It gets read as a real command, so the model goes hunting for one that was never installed, comes up empty & asks you what you meant.

You came for an answer. You got a help desk.

So OP built something genuinely good & pinned the credit on the wrong piece. It works because the words work. The slash is just where he hung the label.

Build Up

None of this makes the list worthless. A good menu is worth keeping.

Think of the Starbucks secret menu. It was never secret, just a list of drinks you wouldn’t have thought to order & people screenshot it for the days they don’t know what they want or feel like trying something new.

Use these the same way. Blank screen, not sure what to ask for? Scroll the list, grab the one that fits, then give the model the context to go with it.

You can even build your own:

Read through everything you know about me and give me 10 commands like /eli5 that fit how I work. Put each on its own line.

Keep the slash if it helps you remember. The model doesn’t need it, but you might.

Original post by u/Obsessivefrugality. Link