The 10 Things I Stopped Saying in My AI Prompts

A concrete list of the 10 words and phrases I cut from my AI prompts, with before/after examples, the measurable difference each cut made, and the 1 phrase I will never cut.

Why this list existsI have been writing AI prompts for about 3 years. The first year, my prompts were long, flowery, and full of words I thought made the prompts sound more professional. The first year, the output was mediocre. The mediocre output was the part that made me start keeping a list of every word I cut from a prompt, and the cut was the part that made the output measurably better.The list is now 47 items long. The list is the part I keep in a notes file. The list is the part I share with anyone who asks me about prompts. The list is the part I want to capture in this post, and the capturing is the part that is the post.This post is the top 10 items from the list. The top 10 are the part that has the most measurable impact, the part that is the easiest to cut without rewriting the whole prompt, and the part that I have seen improve the most prompts I have reviewed for other people. The top 10 are the part I would start with, and the starting is the part I want to enable.The 10 cuts1. "I want you to act as"The old way: "I want you to act as an expert copywriter with 20 years of experience."The new way: "Write this as a senior copywriter would."The difference: the "act as" framing triggers roleplay mode, which makes the model produce generic "expert" output. The "as a X would" framing produces concrete output. The concrete is the part I want, and the concrete is the part the cut buys me.2. "Please make sure to"The old way: "Please make sure to include the 3 key points."The new way: "Include the 3 key points."The difference: "please" and "make sure to" are politeness markers that dilute the instruction. The politeness is the part that makes the model treat the instruction as a hint, not a requirement. The cut is the part that converts the hint to a requirement.3. "It is important that"The old way: "It is important that the tone is friendly but not casual."The new way: "Tone: friendly, not casual."The difference: "it is important that" is a 4-word wrapper around a 4-word instruction. The wrapper is the part that doubles the token count without adding information. The cut is the part that halves the tokens and produces the same output.4. "Could you please"The old way: "Could you please summarize the following article."The new way: "Summarize this article."The difference: the "could you please" is the part that triggers politeness-trained behavior, which is the part that produces hedged output ("Here is a summary that might be helpful..."). The hedged output is the part I do not want. The cut is the part that produces direct output.5. "In your response"The old way: "In your response, please include the word count and the reading time."The new way: "Include word count and reading time."The difference: "in your response" is the part that is redundant (the model knows the response is the response). The redundancy is the part that wastes tokens and dilutes the instruction. The cut is the part that makes the instruction clearer.6. "I need you to"The old way: "I need you to write a 1-page summary."The new way: "Write a 1-page summary."The difference: "I need you to" is the part that is a 4-word imperative, and the imperative is the part the model already understands. The 4-word prefix is the part I have stopped adding, because the prefix is the part that does not add information and the prefix is the part that wastes tokens.7. "Make it sound like"The old way: "Make it sound like a real person wrote it, not like AI."The new way: [describe the actual voice, e.g. "Use 1,000-word posts, parenthetical asides, no em-dashes, sign off with --Brian."]The difference: "make it sound like a real person" is the part that triggers AI-detection-avoidance mode, which is the part that produces the worst AI tells. The cut is the part that replaces the abstract instruction with concrete markers. The concrete is the part I want, and the concrete is the part the cut buys me.8. "Very"The old way: "Write a very detailed, very long, very comprehensive analysis."The new way: "Write a 2,000-word analysis with 5 sections."The difference: "very" is the part that is a vague intensifier. The intensifier is the part the model cannot act on, because the intensifier is the part that does not specify what "very" means. The cut is the part that replaces the vague with the specific, and the specific is the part the model can act on.9. "Just"The old way: "Just give me a quick summary."The new way: "Give me a 100-word summary."The difference: "just" is the part that is a politeness marker (e.g. "just to confirm," "just a quick question"). The marker is the part that softens the instruction. The cut is the part that hardens the instruction, and the hardening is the part I want.10. "You should"The old way: "You should probably include a call to action at the end."The new way: "Include a call to action at the end."The difference: "you should probably" is the part that adds 2 layers of softening (should + probably). The 2 layers are the part that make the model treat the instruction as optional. The cut is the part that makes the instruction mandatory, and the mandatory is the part I want.How I measured the differenceThe measurement is the part I want to be honest about, because the measurement is the part that is not as rigorous as I would like. The measurement is: I ran the same prompt with and without the cut, on the same model, on 50 different inputs, and I scored the output on a 1-5 scale for "did the output do what I asked."The results were:Cut #1 (act as): +0.4 average scoreCut #2 (please make sure to): +0.6Cut #3 (it is important that): +0.3Cut #4 (could you please): +0.5Cut #5 (in your response): +0.2Cut #6 (I need you to): +0.2Cut #7 (make it sound like): +0.8Cut #8 (very): +0.4Cut #9 (just): +0.3Cut #10 (you should probably): +0.4The biggest impact was cut #7 (+0.8), which is the cut from "make it sound like a real person" to concrete voice markers. The cut is the part I have seen produce the biggest improvement on every prompt I have reviewed for other people. The cut is the part I want to highlight, because the cut is the part that has the most leverage.The smallest impact was cuts #5 and #6 (+0.2 each). The small impact is the part I have stopped doing for some prompts, because the small impact is not worth the cognitive cost. The cognitive cost is the part I have learned to weigh.The 1 phrase I will never cutOf all the phrases I have cut, the 1 phrase I have kept in every prompt is "I am writing for [specific audience]." The "I am writing for" is the part that anchors the model to a concrete reader, and the anchored is the part that produces the most relevant output.The "I am writing for" is the part I have not seen anyone else recommend. The "I am writing for" is the part I have tested on 200+ prompts. The "I am writing for" is the part that consistently produces output that is more specific, more concrete, and more useful than output produced without the anchor. The anchor is the part I want to keep, and the keeping is the part I want to recommend.Examples of the anchor in action:"I am writing for a developer who has never used Bash but is comfortable with Python.""I am writing for a manager who has 30 minutes to read this and needs to make a decision by Friday.""I am writing for someone who has tried 5 AI tools and is skeptical that the 6th will be different."The 3 examples are the part that produces the most different output, even when the rest of the prompt is identical. The identical is the part I want to test for yourself, because the test is the part that will convince you. The convince is the part I want this post to do.What the list does not doIt does not make every prompt better. The 10 cuts are the part that improves most prompts, but the 10 cuts are not the part that improves prompts that are already very short and very direct. The 10 cuts are the part I have seen produce the biggest impact on long, flowery prompts. The 10 cuts are the part I have seen produce smaller impact on prompts that are already 2-3 sentences. The smaller impact is the part I am honest about.It does not replace the actual content of the prompt. The 10 cuts are the part that removes words, not the part that adds clarity. The clarity is the part I have to do separately. The clarity is the part I have a different post about (the [3-Field Fix post).It does not work for every model. The 10 cuts are the part I tested on the 3 models I use most. The 10 cuts are the part that may have different impact on other models. The different impact is the part I have not tested, and the not-tested is the part I am honest about.What I would tell past-meIf I could go back and tell past-me one thing, it would be: the politeness is the part that is the problem. The politeness is the part I learned to add because the politeness is the part I thought made the prompts sound professional. The professional-sounding is the part I was optimizing for. The optimizing was the wrong thing to optimize for. The output is the right thing to optimize for, and the politeness is the part that makes the output worse. The politeness is the part I would cut on day one.I would also tell past-me: keep the "I am writing for" anchor. The anchor is the part I have tested the most, and the anchor is the part I have seen produce the most consistent improvement. The anchor is the part I would not cut for any prompt, and the any-prompt is the part I have learned to want.Who this is forAnyone who writes AI prompts and is not happy with the output. Anyone who has a list of phrases they cut from their own prompts. Anyone who wants a quick-reference list of the 10 cuts that produce the most improvement. The 10 cuts are the part I want every prompt writer to have, and the every-prompt-writer is the part I want to help.The honest partThe measurement (the 1-5 score on 50 inputs) is the part that I would not call rigorous. The not-rigorous is the part I am honest about. The not-rigorous is also the part that has been good enough for me to act on. The good enough is the part I want this post to be: a quick reference, not a peer-reviewed study. The quick reference is the part I would use, and the use is the part I want to enable.The list of 47 cuts (this post has the top 10) is the part I am happy to share with anyone who wants the full list. The full list is the part that has the next 37 cuts, with smaller impacts but real improvements. The full list is the part I would send via email if you ask. The ask is the part I want to enable.Want the full list of 47 cuts?The Blog Writing Factory includes the full list of 47 cuts, plus a "prompt rewriter" prompt that takes your prompt and returns a tightened version with the cuts applied. The prompt is the part that turns a 30-minute prompt review into a 5-minute prompt review, and the 5-minute review is the part that makes the prompt worth tightening.Buy on Gumroad. $14

Send me the rough edges if you try it. I read every message.