Most "bad AI answers" aren't the model's fault — they're the prompt's fault. Specifically, they're the fault of a prompt that spends 200 words re-explaining who you are and 12 words on what you actually want. Let's dissect one, line by line, and see exactly where the tokens (and your patience) went to die.
Exhibit A: a real-looking, totally typical prompt
Here's the kind of message that gets pasted into ChatGPT or Claude approximately eleven thousand times a day. Nothing about it is unusual. That's the problem.
"Hi! So a bit of context before I ask my question — I run a small DTC skincare brand called Lumen & Co, we sell serums and moisturizers, mostly to women 25-45, and our brand voice is friendly but science-backed, not overly salesy, we don't like exclamation marks or hype words like 'game-changing' or 'revolutionary.' We're a small team so I wear a lot of hats, I do marketing and some of the customer service too. When you write for us please use a warm but professional tone, keep paragraphs short because our customers read on mobile, and always avoid making medical claims since we're not FDA regulated as a drug. Also we like to end emails with a soft CTA, not pushy. Okay so with all that in mind — can you write me a welcome email for new subscribers to our newsletter? It should probably explain who we are and what they'll get by signing up. Maybe keep it under 200 words. Thanks so much, let me know if you need anything else from me!"
Reads fine, right? Sounds like a nice, thorough person being helpful to the AI. It's also almost 230 words, and roughly 75% of it is going to be retyped again tomorrow, and the day after, and every day this person emails an AI for the rest of the quarter.
The teardown, line by line
"Hi! So a bit of context before I ask my question —"
Verdict: throat-clearing. Zero information. Costs tokens, costs the reader's patience, delays the actual ask by a full sentence. Free to cut, no downside.
"I run a small DTC skincare brand called Lumen & Co, we sell serums and moisturizers, mostly to women 25-45"
Verdict: standing context, re-typed from memory, badly. This is exactly the kind of fact that should live somewhere permanent — not be reconstructed from scratch in a chat box every single time. Worse: what if tomorrow's version of this sentence forgets "mostly" and says "women 25-45" as an absolute? Now the AI has slightly different facts on different days, because a tired human is retyping company info from memory instead of it being defined once and locked in.
"our brand voice is friendly but science-backed, not overly salesy, we don't like exclamation marks or hype words like 'game-changing' or 'revolutionary'"
Verdict: this is the good stuff — and it's being wasted. This is a real, specific, useful style rule. The problem isn't that it's unimportant, it's that it's ephemeral. It exists for one message and then evaporates. Next week, when this same person asks for a product description, they either retype this whole rule (more wasted tokens) or forget to, and get copy full of exclamation marks and the word "revolutionary" that they now have to reject and redo. That back-and-forth is the actual cost of vague, one-off standards — not the first draft, but the second, third, and fourth.
"We're a small team so I wear a lot of hats, I do marketing and some of the customer service too."
Verdict: irrelevant to this task. Interesting biographical detail, changes nothing about how the email should be written. This is the conversational equivalent of a Slack message that starts "so basically" and ends four minutes later at the actual point.
"please use a warm but professional tone, keep paragraphs short because our customers read on mobile, and always avoid making medical claims since we're not FDA regulated as a drug"
Verdict: more standing rules, correctly identified as important, incorrectly placed. "Avoid medical claims" is not a one-off preference — it's a compliance rule that should apply to every single piece of copy this brand ever generates, forever. Burying a permanent legal guardrail inside a single prompt means it only protects you on the days you remember to type it. That's not a style choice, that's a liability sitting in the wrong place.
"we like to end emails with a soft CTA, not pushy"
Verdict: vague, and vague is expensive. "Soft, not pushy" means something different to every reader of this sentence, including the AI. It's going to guess. Maybe it guesses right. Maybe it writes "Explore our bestsellers today!" and you have to explain, again, that "today" reads as pushy to you. Vague instructions don't fail loudly — they fail quietly, by producing a plausible-looking first draft that's just slightly off, which triggers a round of "no, more like—" messages that cost more time than writing the specific instruction would have in the first place.
"can you write me a welcome email for new subscribers to our newsletter? It should probably explain who we are and what they'll get by signing up. Maybe keep it under 200 words."
Verdict: finally, the actual task — buried at word 190. This is genuinely all the AI needed a clear shot at. Everything before it was either throat-clearing, standing context that shouldn't be retyped, or vague style guidance competing for attention with the real request.
"Thanks so much, let me know if you need anything else from me!"
Verdict: nice. Also free to cut. The AI is not going to feel slighted by its absence.
~230 words total. Roughly 30 words are the actual task. Roughly 90 words are standing context that should never need retyping. Roughly 60 words are real style/compliance rules stranded in a one-off message. The rest is filler. That ratio is the whole disease.
What a skill file actually fixes here
A skill file (the system-prompt / Project-instructions document you load once into Claude, ChatGPT, or Gemini) is just a permanent home for everything in that prompt that isn't the task. Brand name, audience, tone rules, banned words, compliance guardrails, formatting preferences — all of it gets defined once, loaded once, and then it's simply true for every conversation after that. You're not retyping standing facts from memory anymore, which means you're not risking a slightly-different version of your own brand voice showing up on a Tuesday because you were rushing.
Load Jake once with your brand voice, tone rules, and banned-word list, and every welcome email, product description, or landing page brief after that starts from a specific, on-brand baseline instead of a guess.
View Jake — Conversion Copywriter →Exhibit B: the "after" version
Assume the skill file above is already loaded — brand name, audience, voice rules, banned words, and the FDA-compliance guardrail are all defined once, permanently. Here's the entire message that gets sent when a real welcome-email task comes up:
"Write a welcome email for new newsletter subscribers, under 200 words, ending with a soft CTA to browse the bestsellers page. Mention what they'll get by signing up (early access to new serum launches, occasional discount codes)."
That's it. Two sentences. No re-explaining the company. No re-explaining the tone. No hoping the AI remembers what "not pushy" means this time, because "not pushy" was already defined in the skill file with an actual example, not a vibe.
Why the after version is cheaper AND better
These are two separate wins, and it's worth pulling them apart:
- Cheaper: the standing context (brand, audience, tone, compliance rules) gets typed once instead of once-per-conversation. Across a week of prompts, that's not a rounding error — it's the difference between a 230-word message and a 30-word one, every single time. Fewer input tokens, fewer minutes spent typing the same paragraph from memory, less mental overhead just getting to the point.
- Better: the AI isn't guessing at unstated standards, because there aren't any unstated standards left. "Soft CTA" isn't a vibe the model has to infer from tone — it's a rule the skill file already spelled out with an example of what soft looks like versus what pushy looks like. That's fewer "no, more like—" follow-ups, which is the real, hidden cost of a vague first prompt: not the first answer, but the three re-writes it takes to get an answer you'd actually send.
Put those two together and the math is stark: a prompt that's 85% shorter and produces a first draft that needs less editing isn't a minor efficiency gain. It's the difference between "AI as a slow, unreliable intern who forgets everything overnight" and "AI as a specialist who already knows how you work."
If your "wasted prompt" is actually a wasted SEO brief — re-explaining your niche, keyword targets, and content rules every time — Serge holds that context permanently so every brief starts at the real task.
View Serge — SEO Specialist →How to run this teardown on your own prompts
Take your last long prompt to ChatGPT or Claude and highlight it in three colors: gray for throat-clearing you'd cut with no loss, yellow for standing facts about you or your company that don't change task to task, and green for the actual task. If yellow is more than a third of the message, that's not a prompt problem — that's a missing skill file. Everything yellow belongs somewhere permanent. Everything green is what you should actually be typing.
In summary:
Most long prompts aren't detailed — they're disorganized, mixing permanent facts about you with the one thing you actually want answered today. Move the permanent stuff into a skill file like Jake (Conversion Copywriter) or Serge (SEO Specialist) once, and every prompt after that gets to be two honest sentences instead of a paragraph of re-introduction.
Plug in how you (or your team) actually use AI chats and get a real monthly token-waste and dollar estimate for re-prompting vs. a persistent skill file. No signup, no email — just the math from this post applied to your numbers.
Try the calculator →Frequently asked questions
What exactly is a "skill file"?
It's a text document — often written in markdown — that you paste into Claude's Project instructions, a ChatGPT custom GPT or Project, or a Gemini Gem, once. It holds everything that should stay true across every conversation: who you are, your standards, your compliance rules, your formatting preferences. After that, every new chat starts already knowing it.
Isn't repeating context in every prompt just "being thorough"?
It feels thorough, but it's actually the opposite of reliable — you're reconstructing the same facts from memory each time, which means small inconsistencies creep in, and you're spending words on setup instead of the actual task. A skill file is more thorough than any prompt, because it's written once, carefully, and doesn't degrade under time pressure.
Will a shorter prompt really produce a better answer?
A shorter prompt on top of a loaded skill file produces a better answer, because the standing rules are still there — they're just not competing for attention with the task inside a single message. A short prompt with no skill file behind it is a different, worse thing: that's just an under-specified request.