Why AI Answers Disappoint: Patterns From Hundreds of Graded Briefs
Published 2026-06-09
Every day we read and grade real briefs that real people write for AI — emails they need to send, raises they want to ask for, decisions they are trying to untangle, things they have to explain to someone who does not share their context. And every day we see the same thing: when an AI answer disappoints, the problem is almost never the model. It is the brief.
That sounds like a blame-the-user take, so let us soften it right away: writing a good brief is genuinely hard, nobody teaches it, and the failure modes are invisible until someone points them out. The good news is that the failure modes are not infinite. After grading hundreds of briefs in our daily challenges, we keep seeing the same five patterns, again and again, across wildly different tasks. Once you can name them, you start catching them in your own prompts before you hit Enter.
Here they are — what each one looks like, why the model stumbles, and how to fix it.
A quick note on how models fill gaps
One piece of plain-English mechanics explains most of what follows, so let us get it out of the way.
A language model is, at its core, a next-token predictor: given everything written so far, it predicts the most plausible continuation. When your brief is specific, "plausible" means "specific to your situation". When your brief is vague, "plausible" collapses to "the most average answer anyone could give to a question like this". The model is not being lazy — it literally has nothing else to work with, so it fills every gap with the statistical middle of all the text it has ever seen.
That is why underspecified briefs produce answers that feel generic, corporate, or weirdly confident about details you never provided. In the worst case, the model invents specifics to fill the void — what practitioners call a hallucination. Most of the fixes below boil down to one move: close the gaps yourself, so the model does not have to guess.
Pattern 1: The goal lives in your head, not in the brief
This is the most common pattern we see, by a wide margin. The writer knows exactly what they want — and forgets to write it down.
What it looks like
Write an email to my manager about the project deadline.
The person who wrote this knows whether they want to push the deadline, defend it, warn about a risk, or quietly shift blame. The model does not. It will produce a polite, content-free status update — technically "an email about the deadline", practically useless.
Why the model fails
"About the deadline" defines a topic, not an outcome. The model predicts the most generic email anyone might write on that topic. There is no goal to optimize for, so nothing in the answer pushes toward one. You read the result and think "that's not what I meant" — and you are right, because what you meant was never in the prompt.
The fix
State the outcome you want the message to produce in the reader, not just the subject matter.
Write an email to my manager. Goal: get the deadline moved from
Friday to next Wednesday without sounding like I'm behind.
Key facts: the design review came back two days late (not my fault),
and I'd rather ship something solid than rush. Tone: calm, confident,
no apologizing. Keep it under 120 words.
Same task, completely different output — because now the model knows what "success" means. This is the core idea behind our clarity principle lesson: a brief is not a description of a document, it is a description of a result.
Pattern 2: No audience, so the AI writes for the generic middle
The second pattern is the invisible reader. The brief describes the task but never says who will read or hear the result.
What it looks like
Explain how a VPN works.
Explain it to whom? A network engineer, a teenager, your CFO, your grandmother? Each of those needs a different vocabulary, a different level of detail, and a different set of analogies. Leave the audience out, and the model aims at an imaginary "average reader" who does not exist — too technical for beginners, too shallow for experts, satisfying for no one.
Why the model fails
Audience is one of the strongest signals a model can use. It changes word choice, sentence length, which details matter and which are noise. Without it, next-token prediction defaults to the middle of the distribution: textbook-flavored prose, mild jargon, hedged everything. The model is averaging over every possible reader because you did not pick one.
The fix
Name the reader and what they already know (or do not).
Explain how a VPN works to my 78-year-old grandmother. She uses
WhatsApp and online banking but doesn't know what an IP address is.
Use one everyday analogy, no technical terms, max 5 sentences.
She mainly wants to know if it keeps her banking safe on hotel Wi-Fi.
This exact task is one of our favorite daily challenges — try explain a VPN to your grandma and see how much the audience line alone changes your score. The same applies to high-stakes audiences: in the ask for a raise challenge, briefs that describe the manager ("numbers person, hates fluff, just survived a budget cut") consistently outperform briefs that only describe the raise.
Pattern 3: Missing constraints, so the model fills gaps with cliches
Even with a clear goal and a named audience, a brief can leave the model too much room. Whatever you do not constrain, the model fills with the most statistically common choice — and the most common choice is, by definition, a cliche.
What it looks like
Write a product description for my handmade ceramic mugs.
You can predict the output without running it: "Elevate your morning routine…", "crafted with love…", "the perfect blend of form and function…". Not because the model is uncreative, but because those are the highest-probability phrases in the sea of product descriptions it learned from.
Why the model fails
Constraints are how you steer a model away from the average. Every fact, boundary, and "do not" you provide prunes the space of plausible continuations. No constraints means the full space stays open, and the center of that space is beige. The model is not ignoring your taste — it has no evidence of your taste.
The fix
Feed it the specifics only you know, and explicitly fence off what you do not want.
Write a product description for my handmade ceramic mugs.
Facts: each one is wheel-thrown in my garage studio in Porto,
glazed in matte sea-green, holds 350 ml, slightly irregular on
purpose. Audience: people who buy from small makers on Etsy.
Voice: warm, a bit dry-humored, first person.
Banned: "elevate", "crafted with love", "perfect blend",
anything that could appear on a mass-market mug.
Length: 80-100 words.
The banned-phrases line feels petty until you try it. It is one of the cheapest, highest-leverage edits we know. One caution: pile in the constraints that matter, but do not paste your entire life story — everything you send competes for space in the model's context window, and a wall of irrelevant detail can bury the three facts that actually steer the answer.
Pattern 4: Asking for everything in one shot
This one is less about wording and more about workflow. The brief tries to get a final, polished, ready-to-ship artifact in a single prompt — and then the writer is disappointed that round one is not perfect.
What it looks like
Write the complete launch plan for my newsletter: positioning,
name, content calendar for 3 months, growth strategy, and the
first three issues. Make it great.
Each of those sub-tasks depends on the answers to the previous ones. Asking for all of them at once forces the model to lock in dozens of decisions you have not weighed in on, then build everything downstream on top of those guesses.
Why the model fails
The model cannot pause and ask "wait — do you want this newsletter to be funny or authoritative?" (well, it can, but a one-shot brief implicitly tells it not to). So it guesses on every fork in the road and compounds those guesses. By the end, the output is a coherent plan for a newsletter that is not quite yours. The longer the chain of unverified guesses, the further the result drifts from your head.
The fix
Ask for a draft to react to, not a deliverable to accept. Treat the first answer as the start of a conversation.
I'm planning a weekly newsletter about urban gardening for
apartment dwellers. Before any plan: give me 3 sharply different
positioning options (one sentence each) and 5 name candidates
per option. I'll pick a direction, then we'll do the calendar.
Three short rounds — direction, then structure, then drafts — almost always beat one giant prompt, because you correct course while corrections are still cheap. We see this in the grading queue constantly: writers who brief in stages get better final artifacts with less total effort. If you want a compact structure for that first round, the 5-line brief template is the format we recommend to everyone.
Pattern 5: No output format, so you get an essay when you needed three bullets
The last pattern is the easiest to fix and somehow the most persistent. The brief says what to produce but not what shape it should take — so the model defaults to its favorite shape: a well-organized essay with an introduction, several balanced paragraphs, and a tidy conclusion.
What it looks like
Compare these two job offers for me. [details of offer A and B]
You wanted a decision aid: a short table, the two or three factors that actually differ, and a recommendation. You got 600 words of "On the one hand… on the other hand… ultimately, the choice depends on your personal priorities." Thanks.
Why the model fails
Essay-shaped answers dominate the model's training data for open questions, so essay is the default prediction. Format is also where models are most obedient — they follow explicit structure instructions remarkably well — which makes this gap especially wasteful. You are one sentence away from the output you wanted, and that sentence is missing.
The fix
Describe the artifact: its skeleton, its length, even its labels.
Compare these two job offers. Output format:
1. A table: salary, equity, commute, growth, risk — one row each,
with a one-line note per cell.
2. "What actually differs": max 3 bullets.
3. "Recommendation": one sentence, pick one, no hedging.
No introduction, no conclusion, nothing else.
"No introduction, no conclusion" alone removes half the filler. If you want to go deeper — tables, JSON, fill-in-the-blank templates, length budgets — our output formatting lesson covers the patterns that work and the ones that quietly backfire.
The five patterns, side by side
When a brief earns a low score in our game, the feedback almost always points at one of these:
- Goal in your head — the brief names a topic, not an outcome. Fix: write the result you want, not the document.
- No audience — the model writes for the average of everyone. Fix: name the reader and what they know.
- No constraints — gaps get filled with cliches. Fix: add your specifics, ban what you hate.
- One-shot everything — compounding guesses. Fix: stage it; react to a draft.
- No format — essay by default. Fix: describe the artifact's shape and length.
Notice what is not on the list: magic words, secret jailbreak phrases, "act as a world-class expert" incantations. In our grading we see well-briefed prompts with zero tricks beat trick-laden vague ones every single time. Briefing is a communication skill, and it transfers — the habits that earn five stars in the game are the same habits that make you clearer with colleagues.
Practice beats theory
Reading about the five patterns is the easy part; catching them in your own writing is where the skill actually forms. That is exactly what CUCKCO.DE is for: one short, real-life briefing challenge a day — an email, a negotiation, an explanation, a decision — scored in seconds by an AI coach that tells you which of these patterns tripped you up and how to fix it. It is free, it takes a few minutes, and the improvement curve is genuinely fun to watch.
Try today's challenge — and the next time an AI answer disappoints you, you will know exactly which line of your brief to fix.