Every LLM defaults to the same failure mode: it answers your three-sentence question with four paragraphs, hedges every claim into mush, and buries the one thing you actually needed in the middle of paragraph three. Then you skim it, miss the important part, and either re-ask the question or act on the wrong takeaway.
This isn’t a model problem. It’s a prompting problem wearing a model’s clothing.
I pay for these tools out of my own pocket, and I’ve watched myself do this: ask a vague question, get a vague-shaped answer back, and blame the AI for being unhelpful. The AI didn’t fail. I never told it what “helpful” looked like.
Two things fix this. One is on you — sharpen your reading and critical thinking so you can tell a good answer from a padded one. The other is on your prompt — structure and constraints that make padding structurally impossible. This piece is about the second one, with the specific phrases that work and the scenarios where each one earns its keep.
Why LLMs ramble in the first place
Three reasons, and none of them are malice:
- They’re trained to be thorough, not efficient. Completeness reads as helpfulness in training data. Nobody rates a one-line answer five stars unless the prompt asked for one.
- They hedge because the world is uncertain and they know it. “It depends” is often true. It’s also useless if you needed a decision by 3pm.
- You didn’t tell them to stop. Left unconstrained, a model will fill the space you gave it — and by default, that space is “as much as seems reasonable,” which is almost always more than you wanted.
The fix isn’t a smarter model. It’s a sharper prompt. Here’s the full toolkit. The one-page version is the verbosity cheatsheet (PDF).
Structure prompts: force the shape before the content
The single highest-leverage move is specifying the shape of the answer before the model starts generating. A shape constrains the content — you can’t ramble inside a 3-bullet limit.
| Prompt | Use it when |
|---|---|
| Return this as a numbered list, one line each. | You need a checklist, not an essay. It forces one item per line instead of prose that hides where one idea ends. |
| Give me a hierarchical bulleted list — main points, with sub-bullets for detail. | The topic has real categories, and you want that structure visible. |
| Numbered list of items, each with a 2–3 sentence explanation. | You want more than a label and less than a full section. |
| Format as a table: [columns you want]. | You are comparing options. A cell has no room for a hedge. |
| TL;DR first, then the details. | You want the payload immediately, and the reasoning only if you keep reading. |
| Give me your answer, then your reasoning — not the other way around. | The model thinks out loud before it commits. Flip that order. |
| Problem, then solution, then example. Nothing else. | You want those three sections and not a fourth. |
| Give me a decision, not a menu of options. | You want a call, not “you could do A, or B, or C.” |
Tone and audience prompts: calibrate the vocabulary, not just the length
Length and complexity aren’t the same lever. A short answer can still be dense; a long one can still be simple. These prompts control density.
| Prompt | Use it when |
|---|---|
| Explain like I’m a 6th grader, a 10th grader, or someone with a PhD in the field. | You need a vocabulary target. A grade level beats “explain simply.” |
| Use simple language, no jargon. | You are handing this to someone else and do not have time to translate it. |
| Explain like I’m a busy executive with 30 seconds. | The audience is impatient, not uninformed. |
| Write like a text message to a colleague, not a report. | You want the answer in the register you would actually use. |
| No corporate jargon or buzzwords. | The model is reaching for “leverage,” “synergy,” and “unlock.” |
| Assume I already know the basics of X — skip the intro. | You are tired of the same preamble before every answer. |
Length and constraint prompts: cap it, don’t just ask nicely
“Be concise” is a request. A number is a constraint. Constraints work better.
| Prompt | Use it when |
|---|---|
| Under 100 words. | “Concise” has not been working. A number the model can check beats a vibe. |
| One paragraph, max. | You want prose, and you still want a ceiling. |
| Answer in a single sentence if at all possible. | You suspect the honest answer is short, and you want to see if the model was only padding. |
| Give me only what changed — don’t recap what we already covered. | You are several turns in, and every reply restates the conversation. |
Scenarios that deserve their own prompt — and why
Beyond format and length, there are recurring failure patterns in LLM output that each need a targeted counter-prompt. These are less about formatting and more about content behavior.
| Problem | What goes wrong | Prompt |
|---|---|---|
| Hedging | It qualifies everything into “it depends,” which is often true and useless when you need to act. | Give me your single best answer, then briefly note the main exception. |
| Yes-man | A critique still finds a way to validate the plan, and it looks like feedback. | Push back if you disagree. Argue the strongest case against this before you tell me it’s a good idea. |
| Generic advice | The advice would fit anyone in the category. The more specific the constraints, the less room for a generic pattern. | Give me advice specific to my team size, budget, deadline, and tools, not general best practices. |
| Buried lede | The recommendation is in the last sentence of paragraph four. | Put your recommendation in the first sentence. Context after, if needed. |
| Re-explaining | Every follow-up restates context you both already have. Set this once per conversation. | Just answer the new question. Skip the recap. |
| Over-commented code | Every line gets a comment, including i = i + 1. |
Just the code. Comment only the non-obvious parts. |
| False balance | Both sides are presented as equal because balance feels safer. Skip this prompt when the question is genuinely contested. | Where does the evidence actually point? Don’t false-balance this. |
| Scope creep | A narrow question gets an answer to a broader one. | Answer only what I asked. If something important is missing, flag it in one line at the end. |
| Everything is important | A summary gives every point the same weight. | Rank these by importance, don’t just list them in the order they occurred to you. |
The part that isn’t about prompting at all
None of this replaces judgment. A well-structured, appropriately-toned, tightly-scoped answer can still be wrong — just wrong in a more readable format. Prompting technique controls the packaging. It doesn’t substitute for reading critically, checking claims against what you actually know, and noticing when a confident-sounding answer is confidently sounding rather than confidently correct.
The skill stack, in order: read critically first, prompt precisely second. Precise prompts just mean you’re spending your critical-thinking budget on the content instead of on excavating it from the padding.