“Improve this text.” The AI hands back corporate mush. “No, better.” More mush, different tie. “PUNCHIER!” By message ten you’re typing in capitals at a machine that keeps apologizing. Sound familiar?
Developers have a word for this kind of trap: an antipattern: a solution that looks sensible, that everybody reaches for, and that reliably backfires. Good recipes get names so you can copy them (we’ve catalogued those in the patterns section); the traps get names so you can catch yourself mid-mistake. Here are the seven we see everywhere in prompting, and, each time, the fix. You’ll see: it’s not rocket science.
One mental image for the whole article, the same one as in our prompt fundamentals: a prompt is a brief to a brilliant freelancer: talented, absurdly fast, and paid by the word exchanged. The seven antipatterns below are the seven clients every freelancer dreads.
1. The crystal ball
“Make me something good.” Then you judge the result — harshly.
Here’s the mechanism: a model completes your text with the most plausible continuation. Give it a vague ask, and the most plausible answer to a vague ask is the statistical average of everything: that generic, faintly corporate tone you recognized instantly. The mush isn’t a malfunction; it’s exactly what “improve this text” means, on average, across the whole internet.
The fix is the four-part brief from prompt fundamentals: a role, the context, one precise task, the expected format. “Rewrite this for the newsletter: subscribers are non-technical, keep it under 120 words, end with a question” gets you in one pass what “improve this” won’t get you in ten.
2. The kitchen sink
The opposite client: the 40-page brief. Five roles, thirty rules, three old versions “for reference”, every instruction ever written, pasted into every request, just in case.
Two things go wrong. First, rules get dropped: research on long contexts showed models exploit the start and the end of a prompt much better than what sits in between; the phenomenon has a name, lost in the middle, and we saw it in the context-window article. Rule 17 of 30, buried mid-prompt, is the one the model forgets. Second, you pay for the pile on every single message, because the whole brief is re-sent as input tokens each turn.
The fix: everything needed, nothing more. Promote the rules that are always true into custom instructions (written once, applied everywhere), and keep the prompt for what’s specific to this task, with the crucial constraints at the top or the bottom, not drowned in the middle.
3. The pink elephant
“Whatever you do, do NOT mention our competitor.” Guess who shows up in paragraph one.
It’s the oldest trick in psychology: to not think about a pink elephant, you first have to think about it. Models have the same weakness, for a mechanical reason: the forbidden word is now sitting in the context, radiating probability onto every next token. Negations are also simply fragile: “don’t be formal” is the kind of instruction a model drops first, as we noted in prompt fundamentals.
The fix: name the target, not the void — and give an escape route. Instead of “don’t mention competitors”: “Talk only about our product. If a comparison is unavoidable, point to the comparison page instead.” Positive instructions give probability something to flow toward.
4. The adjective catalog
“Professional but warm. Punchy but not aggressive. Modern but timeless.” A human freelancer nods politely and understands nothing; the model does the same, minus the politeness.
Adjectives are cheap because they’re ambiguous: your “punchy” and the internet’s average “punchy” are different things. There’s a tool built precisely for this: few-shot. Show two or three real examples of the tone you want, and the model copies the pattern; models are far better at imitating than at decoding vocabulary, as the few-shot pattern details.
The fix: for anything stylistic, one good example beats ten adjectives. An example costs a few hundred input tokens; with prompt caching, pennies. Ten adjectives cost three rounds of “no, not like that”.
5. The zombie thread
Message 63 of the same conversation. You’ve corrected course twelve times, pasted three versions of the document, changed your mind twice about the format. And the answers keep getting… worse?
Two reasons. Quality: every correction, including the overruled ones, is still in the context: v1 of the doc sits next to v3, “make it shorter” next to “actually, longer”, and the model averages over the contradictions. Cost: each new message re-sends the entire history. A thread whose history has grown to 20,000 tokens bills you 20,000 input tokens per message. Ten more exchanges, 200,000 extra tokens. (Prompt caching softens the bill by up to ~90% on the stable prefix; it does nothing for the confusion.)
The fix: harvest and restart. Pull the two or three decisions that emerged from the thread (“audience: beginners, format: table, tone: like this example”), fold them into a fresh, upgraded prompt, and open a new conversation. A thread is a workshop, not a home; the context-window article explains why the model itself works this way.
6. The magic words
“You are the world’s #1 expert.” “I’ll tip $200.” “My career depends on this.” “Take a deep breath.”
Time for some honesty, because this one isn’t pure superstition: in a 2023 Google DeepMind study, the prompt “Take a deep breath and work on this problem step-by-step” genuinely beat every other tested instruction on a math benchmark — for that model, that year. Community tests on tipping and threats found small effects too (on some models, sometimes). That’s exactly the problem: these effects are model-specific, unstable, and fading with each generation, while the boring stuff (context, examples, format, one clear task) pays on every model, every time. An incantation carries zero information about your task; a constraint carries plenty.
The fix: swap flattery for facts. “You are the world’s best copywriter” → “Here are two past newsletters; match their tone.” And if you’re attached to a lucky charm, fine, but measure it (see antipattern 7) instead of believing it.
7. “It worked yesterday”
You tweaked the prompt on one example, it looked great, you shipped it. Three other cases silently broke.
A prompt sitting inside an app is code: it has inputs, outputs, and regressions. Nobody ships code after running one manual test; yet prompts get edited “on vibes” daily. The result is whack-a-mole: each fix breaks something you’re not looking at.
The fix has a name, evals: a fixed set of real cases with expected results, re-scored after every prompt change. Twenty rows in a spreadsheet is a perfectly honest start; we wrote a whole article on it, and a pattern with the checklist.
Honorable mentions
Three more, one line each: using a chatbot as a search engine (that’s hallucination territory: ground it, RAG exists for this); reaching for the temperature dial to fix what is a prompt problem (the mixing console won’t save a bad brief); and the mega-prompt doing five jobs at once, when chaining small steps is cheaper and easier to debug.
The recap table
| # | Antipattern | The tell | The fix |
|---|---|---|---|
| 1 | The crystal ball | “make it good”, then disappointment | role + context + one task + format |
| 2 | The kitchen sink | 30 rules “just in case”, every time | everything needed, nothing more; permanent rules → instructions |
| 3 | The pink elephant | a brief made of “don’t”s | name the target, give an out |
| 4 | The adjective catalog | “punchy but classy, modern but timeless” | show 2–3 examples (few-shot) |
| 5 | The zombie thread | correction #12, same conversation | harvest the decisions, restart fresh |
| 6 | The magic words | “$200 tip”, “world’s best expert” | swap flattery for constraints; measure |
| 7 | “It worked yesterday” | tested once, shipped | evals: 20 real cases beat vibes |
The common root
Look at the seven again: they all treat the model as a mind-reader (1, 4), a psychology subject (3, 6), a colleague who remembers everything (5), or a slot machine (7) — anything but what it is: a very literal, very fast freelancer, paid by the token. Brief it like the professional it imitates, and most of these evaporate.
That’s the whole point of naming antipatterns: next time an answer disappoints, don’t type “better!!” — ask which of the seven was that, and the fix follows automatically. See? It’s not rocket science.
