I have spent the last eighteen months testing AI tools every day for real work, including writing entire books with Claude and rebuilding most of my workflow around what these models can and cannot do, and the single most useful thing I can give you in a free PDF is the small handful of tricks that produced an actual “wait, what” moment when I tried them. Not the demos that sound impressive on a YouTube thumbnail. The ones I still use weeks later because they keep paying off.
This is not a list of fifty prompts you will skim once and forget. It is five tricks and five prompts, all tested with the free tiers of ChatGPT and Claude, and each one is here for the same reason: it changed something I was doing every day.
You will need about twenty minutes to read this. You will need maybe two minutes to test the first trick. If it does not save you time on the next thing you do, ignore the rest and unsubscribe with my blessing, because I would rather you read nothing than read something that wasted your evening.
How to use this PDF
Each trick has a copy-paste prompt and one short note explaining why it works. Read the why. Tools change every month and the principle behind each trick is what will still be useful when ChatGPT 6 or Claude Sonnet 5 arrive. Skip the principle and you will be stuck the moment something updates.
The prompts assume you are using a current free-tier model (ChatGPT, Claude, or Gemini). They work in all three. Where the behavior differs, I will tell you.
PART 1: THE 5 TRICKS
Trick 1: Voice cloning from a writing sample
What it does: Gets the AI to write in any voice, including yours, your boss’s, or an author you admire, by showing instead of telling.
The thing nobody tells you about prompting an AI to “write in my voice” is that describing a voice does not work. I wasted a full week writing elaborate descriptions of my own writing style and getting back text that sounded like a marketing brochure pretending to be me. The fix turned out to be obvious in hindsight: the AI does not need your analysis of the voice, it needs the voice itself.
Copy this prompt:
You are writing as [name or "the author of the sample below"].
Here is a sample of their writing. Read it carefully and notice the
sentence length, the rhythm, the word choices, and the way they
construct paragraphs. Do not analyze it. Just absorb it.
[Paste 400 to 600 words of their actual writing here]
Now write a 300-word section about [your topic] in this exact voice.
Match the sentence length, the rhythm, and the word choices. Do not
add formality that is not in the sample. Do not summarize at the end.
Stop when the section is done.
What you will see: Output that genuinely sounds like the source. Not perfectly. The AI will still default to certain tics that survive any voice instruction, like adding a closing summary sentence or smoothing out anything jagged. But on the first pass you will get 80% of the way there, which is the difference between text you can edit and text you have to throw away.
Why this works: Pattern matching is what these models are good at. When you describe a voice, you give them an abstraction and they match the abstraction, which is why you keep getting “professional yet approachable” output. When you give them the writing itself, they match the writing. Always show, never describe.
Trick 2: Screenshot interrogation
What it does: Turns any image into something you can ask questions about, including charts, dashboards, contracts, handwritten notes, and printed documents.
Last month I had to summarize a board deck that someone sent me as a screenshot collection because they did not have the original PowerPoint. Twelve images. In the old world, I would have retyped the key numbers into a spreadsheet and built the summary by hand, which is roughly an hour of work I do not enjoy. Instead, I dragged all twelve images into Claude and asked one question: “Summarize the financial story across these slides and flag any number that looks inconsistent between slides.”
It found two inconsistencies in 30 seconds. One was a typo. The other was an actual reporting error that the team had not noticed.
Copy this prompt:
I am attaching one or more images. Look at them carefully and do
the following:
1. Tell me what type of document or chart this is
2. Pull out the key numbers, dates, and names
3. Flag anything that looks inconsistent or surprising
4. Give me a 5-bullet summary of what someone would need to know
If something is unclear in the image, say so instead of guessing.
What you will see: The AI will read text it can see, transcribe numbers from tables, describe charts in terms of what they actually show, and tell you when something is illegible. The “say so instead of guessing” line is critical because without it, the model will fill in details that are not actually visible.
Why this works: Modern multimodal models can see images about as well as a human reading at speed. They are not magic and they will misread handwriting or low-resolution text, but for any printed or screen-captured content, the accuracy is high enough to trust with a verification pass. The bottleneck is no longer the AI’s ability to see. It is your ability to ask the right question about what it sees.
Trick 3: Ask the AI to write the prompt
What it does: Outsources the hardest part of using AI, which is figuring out the right way to ask, to the AI itself.
This one made me feel slightly stupid the first time I tried it. I had been struggling for ten minutes to phrase a request for a marketing brief that would actually produce a usable output. Then I stopped, opened a new chat, and typed: “I want to write a marketing brief for a new product launch. Write me a prompt I can use that will produce the best possible brief, including all the context I should provide and the structure I should ask for.”
Claude wrote me a prompt that was better than anything I had been about to type. I copied it, pasted it into a new chat, filled in my product details, and got a brief that needed almost no editing. The whole exercise took six minutes including the meta-prompt.
Copy this prompt:
I want to use AI to [your task]. Before I write my actual prompt,
help me write a better one.
Write a complete prompt I can copy and paste that will produce the
best possible result for this task. The prompt should:
- Include the role the AI should take
- Specify what context I need to provide
- Define the output format I should ask for
- Anticipate the most common mistakes and instruct against them
- Include one example of what good output looks like, if helpful
After the prompt, briefly explain why you structured it this way so
I can adapt it for similar tasks in the future.
What you will see: A multi-paragraph prompt with placeholders for your actual content, plus a short explanation of the design choices. You will use the prompt with another chat session for the actual work.
Why this works: AI models have read more bad prompts and good prompts than any human ever will. They have a strong implicit sense of what produces good results because they can predict the kind of output a given prompt structure will lead to. When you ask them to design the prompt, you are using their pattern-matching strength to bypass your own weakness, which is the blank page.
Trick 4: The opposite-of-what-you-believe argument
What it does: Finds the holes in your thinking before someone else does.
I run a publishing business and I make pricing, positioning, and product decisions every week. Some of those decisions are bad and I would like to know which ones before they cost me money. The problem is that I am not the right person to find the flaws in my own reasoning, because I already convinced myself the reasoning was good or I would not be doing it.
The trick I use now: paste my decision and reasoning into Claude, and ask it to argue the opposite as convincingly as possible. Not as a polite “have you considered” exercise. As a real argument.
Copy this prompt:
I am about to make this decision: [describe your decision]
My reasoning is: [list your reasons in 3 to 5 bullet points]
Take the opposite position. Argue, as convincingly and specifically
as you can, why this decision is wrong. Steelman the opposing case.
Cite evidence, name failure modes, and identify the strongest
counterexample to my reasoning. Be direct. I am not looking for
balanced perspective. I am looking for the strongest possible
attack on my own thinking so I can stress-test it.
What you will see: A pointed argument against your decision with specific failure modes you might not have considered. About a third of the time, the AI will surface a real risk you missed. About a third of the time, it will raise something you already thought about, which confirms you were thorough. The remaining third will be weak counterarguments that you can dismiss, which is also useful information.
Why this works: Steelmanning is a known technique in serious thinking, but most humans are bad at it because they cannot mute their own beliefs long enough to argue the other side honestly. AI has no skin in the game on your decision and will happily produce the strongest version of the opposing case if you ask for it. The “be direct” and “not looking for balanced perspective” lines are necessary because the default mode is to hedge, and you do not want hedged criticism, you want sharp criticism.
Trick 5: Document Q&A on long PDFs
What it does: Turns a 50-page document into something you can have a conversation with instead of reading cover to cover.
The first time I did this with a real document, I had a 90-page service agreement to review before signing. I uploaded the PDF to Claude and asked: “What are the three most aggressive clauses in this contract from the customer’s perspective, and what would I need to negotiate to fix each one?”
Forty seconds later, I had a list of three clauses with page references, the specific language that was problematic, and suggested edits. I still read the contract myself, because I am not handing legal review to an AI, but the AI’s list pointed me directly at the clauses that mattered. Reading time went from two hours to forty minutes, and I caught one additional issue the AI missed by reading carefully on top of its summary.
Copy this prompt:
I am attaching a [contract / report / research paper / book chapter
/ etc.]. I need to understand it for [your purpose].
Do the following in this order:
1. Tell me in 3 sentences what this document is and who wrote it
2. List the 5 most important things in it for someone with my purpose
3. Flag anything that looks unusual, aggressive, or inconsistent
4. Tell me what is NOT in this document that I might expect to find
Use page or section references for everything you mention so I can
verify it myself. If the document is unclear in places, say which
sections need a careful human read.
What you will see: A structured summary that points you at the parts of the document that actually matter for your purpose, with references you can check. The “what is not in this document” line is the most underused part of this prompt and often produces the most valuable output. Missing information is harder to notice than misleading information, and the AI is good at catching it.
Why this works: Long documents are mostly skippable for any specific purpose. A contract has maybe ten clauses that matter for you and forty that are boilerplate. A research paper has one or two findings that affect your decision and twenty pages of methodology that do not. Asking the AI to filter for your specific purpose collapses the time cost of the document to a fraction without losing the parts you actually needed.
PART 2: THE 5 PROMPTS
These are copy-paste ready. Save them somewhere you can find them again. They are the prompts I use most often for repeat tasks where I do not want to think about the prompt itself.
Prompt 1: The Decision Matrix Builder
Use this when: You have 3 or more options, multiple criteria, and you keep going in circles.
I am choosing between these options: [list 3 to 6 options]
The criteria that matter to me are: [list 4 to 6 criteria]
Build me a decision matrix that:
1. Scores each option on each criterion (1 to 5)
2. Lets me weight the criteria (assume equal weight unless I say
otherwise. I will adjust)
3. Calculates a weighted total for each option
4. Names the option that wins by the math
5. Names the option that wins on instinct, based on the strongest
single criterion
If two options are within 10% of each other on the weighted score,
say so and tell me what tiebreaker would matter most.
One tip: Most decisions are not actually close. The matrix will usually confirm what you already suspected, which is fine, sometimes you just need the confirmation to commit. When the matrix surprises you, that is the high-value output.
Prompt 2: The Email Compressor
Use this when: You wrote a long email and you can feel that it is too long but you cannot see what to cut.
Cut this email down to half its length without losing the meaning,
the relationship cues, or the request. Keep my voice. Do not add
anything I did not say. Do not summarize at the end.
Show me the cut version, then in one paragraph below, tell me what
you removed and why.
Email below:
[paste your email]
One tip: The “tell me what you removed” line is what makes this useful instead of just shorter. You will see your own filler patterns, and after a few rounds you will start writing tighter the first time.
Prompt 3: The Honest Read
Use this when: You wrote something and need an honest read before you send or publish it.
Read what I wrote and tell me what is weakest about it. I am not
asking for a persona or a roleplay. I am asking for these specific
checks:
- Where does the argument get muddled
- Which sentences could be cut without loss
- What claim is unsupported or feels overstated
- Where am I burying the most interesting point
- What would a skeptical reader push back on hardest
Be direct. I am not looking for encouragement. I am looking for the
problems so I can fix them.
Text below:
[paste your text]
One tip: Run this on your draft before you run any “polish” or “improve” prompt. The improvement prompts smooth surfaces. This one shows you the structural problems that surface polishing will hide instead of fix.
Why this prompt has no persona. Most prompt examples for this kind of task open with “You are a sharp, experienced editor” or “Act as a Pulitzer-winning critic.” That works, but it also asks the model to roleplay a person it is not. I do not write that way and I do not want to teach anyone to. The checks above do the same job without the costume.
Prompt 4: The Meeting Prep Brief
Use this when: You have a meeting in an hour and you want to walk in actually prepared instead of vaguely prepared.
I have a meeting about [topic] with [people and their roles]. The
purpose of the meeting is [your purpose]. The desired outcome is
[your outcome].
Build me a 1-page prep brief that includes:
1. The single most important thing I should remember to say
2. Three likely objections or concerns and how I would respond to
each in 1 to 2 sentences
3. Two questions I should ask to better understand their position
4. The "graceful exit", if the meeting goes sideways, what is the
minimum acceptable outcome that still moves things forward
Keep it tight. I will read this in the elevator.
One tip: The “graceful exit” line is the part most prep templates skip. Knowing your minimum acceptable outcome before you walk in is what keeps you from giving away more than you should when the conversation does not go the way you hoped.
Prompt 5: The Brain Dump Organizer
Use this when: You have a head full of unsorted thoughts and you need to turn them into something you can act on.
I am about to dump everything I am thinking about [topic / project /
problem] into the next message. It will be unstructured. Do not
respond yet, just acknowledge and wait for the dump.
When I send the dump, do this:
1. Group the thoughts into 3 to 6 themes
2. For each theme, identify the underlying concern, question, or
decision
3. List the specific actions hidden in the dump (things I can
actually do, not things I am worried about)
4. Flag anything I said that contradicts something else I said
5. Tell me what the dump suggests I should focus on first
Do not add advice I did not ask for. Just organize what is there.
One tip: Send the prompt first, wait for the acknowledgment, then send the dump. This produces noticeably better organization than trying to do both in one message because the model has the structure in mind before it sees the chaos.
The 3 rules that make any prompt better
If you only remember three things from this PDF, remember these.
Rule 1: Specific beats clever. The single biggest improvement most people can make to their prompts is to add concrete details. “Write a marketing email” is bad. “Write a 200-word marketing email to existing customers introducing our new pricing tier, opening with a question, no exclamation points, and ending with a soft CTA” is good. The model is not psychic. It cannot read your mind, your brand guide, or your audience. Tell it.
Rule 2: Show before you describe. When you want a specific style, tone, or format, paste an example of what good looks like. Examples are worth a hundred adjectives. This is the principle behind Trick 1 and it applies to every prompt you will ever write.
Rule 3: Ask for the why. When the AI gives you something good, ask it why it made the choices it made. When it gives you something bad, ask it why it interpreted your prompt the way it did. The why is where the learning lives, and after a month of asking why, you will be writing prompts that work the first time.
What to do next
If any of this saved you time today, you will get more of it. I send The AI Stash newsletter whenever I have something tested and worth your inbox: a trick that worked, a tool I tested, a thing I would skip. No hype, no game-changers, no rapidly evolving landscapes. Just what is actually working for me right now and what is not.
Sign up at theaistash.com/newsletter if you want it. Do not sign up if you do not. The unsubscribe link works on the first email.
If you have a prompt or a workflow that has saved you serious time, hit reply on any newsletter email and tell me about it. The best ones end up tested and written about, with credit. The worst ones still teach me something.
That is everything I planned to give you. Now go test the first trick on something real, before you forget you read this.
- Lauri
Tested: April 2026 with ChatGPT (free tier and Plus), Claude (free tier and Pro), and Gemini (free tier). The five tricks worked across all three. The five prompts worked across all three. Any model from late 2025 onward will handle these. Earlier models may struggle with the longer prompts.
Not tested: every possible edge case. If a prompt does not work for your specific use, hit reply and tell me what happened.
TheAIStash.com, built for people who want to use AI without the hype.