Open a new chat. Type "Act as our brand's copywriter. Here's our tone: confident but not salesy, no exclamation points, always speak to the reader as 'you.' Here's our audience: mid-market SaaS buyers who are skeptical of hype. Here's our format: short paragraphs, no em dashes, headline under 60 characters..." Paste the three-paragraph brand guide you keep in a Google Doc. Now, finally, ask your actual question.
If that paragraph made you wince a little, you already know the problem. Most people who use Claude or ChatGPT for real, repeated work — marketers, support leads, analysts, founders — do some version of this every single time they start a new conversation. It feels like a two-minute tax. It is not a two-minute tax. It is a recurring cost with three separate price tags: tokens, time, and consistency. Let's add them up.
The token cost: paying to re-explain yourself
A typical "context dump" — brand voice, audience, tone rules, formatting preferences, a few do's and don'ts — runs somewhere in the neighborhood of 300 to 600 tokens once you count the instructions themselves plus the natural throat-clearing around them ("just to remind you," "like we discussed," "as always"). That's before the actual question gets asked. It's pure overhead, paid every single time, because a fresh chat has no memory of what you told it yesterday.
Now multiply. A marketer or copywriter who starts, say, 8-10 fresh chats a week to write emails, ad copy, landing page variants, and social posts is re-sending that same context 400+ times a year. At 300-600 tokens per re-explanation, that's somewhere between 120,000 and 240,000 tokens a year spent on nothing but re-establishing who you are and how you want things written — tokens that produce zero new output.
Here's the thing worth being honest about: at current API pricing, a few hundred thousand tokens of plain input text is not going to bankrupt anyone. Even at a few dollars per million tokens — and pricing varies by model and shifts over time, so treat any specific number as directional, not gospel — the raw context-repetition cost for one person is usually a matter of a few dollars a month, not hundreds. The token cost alone rarely justifies the fix on its own. But it's the smallest of the three costs, and it compounds the moment you're not the only person on the team doing this.
A prompt pack that builds a real brand voice document from your actual writing samples — not vague adjectives — so you can paste it once and stop reconstructing it from memory every session.
View AI Brand Voice Guide →The time cost: the part nobody puts on a spreadsheet
Tokens are cheap. Minutes are not. Think about what actually happens when you start a fresh chat for real work: you have to remember what you told the model last time (or dig up the doc where you wrote it down), you paste or retype it, you tweak the wording because you're never quite consistent with yourself, and then — only then — do you get to the thing you actually opened the chat to do.
Call it a conservative 90 seconds to 3 minutes per session, depending on how much context the task needs. For someone running 8-10 fresh chats a week, that's 15-45 minutes a week gone before any real writing happens. Over a year, that's somewhere between 13 and 39 hours — the better part of a full work week — spent typing the same setup instructions into a text box, over and over, for no reason other than the AI forgot everything the moment the last chat closed.
And that estimate is generous. It assumes you remember the context correctly and don't have to go hunting for the brand doc, the tone guide, or the "how we write headlines" note buried in a shared drive. Add that search time and the real number climbs fast.
The consistency cost: the one that actually hurts
This is the cost most people underweight, and it's the one that shows up in the output, not the invoice. When you retype your context from memory every session, you don't retype it identically. One day you say "confident but not salesy." The next day, rushing, you skip the tone note entirely and just ask for "punchy copy." A week later you remember to mention the audience but forget the formatting rule about em dashes.
The AI has no way to know which version of "you" is the real one — it just does its best with whatever partial context lands in that particular chat. The result is copy that drifts: sometimes on-brand, sometimes generic, sometimes needing a full rewrite because the tone missed. If you've ever compared two pieces of AI-drafted copy from the same person and wondered why one nails the voice and the other reads like a stranger wrote it, this is almost always why — not a worse prompt, just a differently-remembered one.
That inconsistency has a cost too, even if it never shows up as a line item: more editing rounds, more "can you try again," more copy that quietly doesn't match the brand and gets published anyway because nobody caught it. Unlike the token cost, this one scales badly — the more people on a team improvising their own version of "our brand voice" into a chat box, the more the brand itself starts to drift.
A senior conversion copywriter persona that researches before it writes — VOC mining, messaging hierarchy, objection mapping — configured once and identical in every chat, no re-briefing required.
View Jake — Conversion Copywriter →Why "just save it in a doc" doesn't actually fix this
The obvious rebuttal is: keep the brand guide in a doc and paste it in every time. Plenty of people do exactly that. But pasting isn't free either — it's still the same tokens, still most of the time cost (you still have to find the doc, open it, copy it, paste it, and often trim it to fit), and it still leaves room for using an outdated version of the doc, or pasting the wrong section, or forgetting the doc even exists three months from now when someone new joins the team.
A saved doc is a step better than retyping from memory. It is not the same as a system that already knows who it's supposed to be the moment you open the chat.
What a skill file actually changes
A skill file — a structured system prompt you load once as a project instruction or custom instruction in Claude, ChatGPT, or Gemini — moves all of this from "re-explained every session" to "configured once, reused forever." You write the brand voice, the audience, the tone rules, the format preferences, and the role the AI should play a single time, with the care of writing a real brief instead of a rushed chat message. After that, every new chat starts already knowing it.
- Tokens: the context lives in the system layer, not retyped as fresh input every session.
- Time: no searching for the doc, no retyping, no tweaking — you open the chat and start on the actual task.
- Consistency: the AI reads the exact same instructions every time, so the output doesn't drift depending on how rushed you were when you set up that day's chat.
For a content team specifically, this is where it compounds fastest. A content marketing manager isn't just writing one piece — they're running an editorial calendar, briefing writers, and keeping dozens of pieces on-brand across weeks. Re-explaining the content strategy and brand system every time doesn't just cost that one person minutes; it means every piece of content is only as consistent as whoever happened to remember the full brief that day.
A senior content marketing manager persona that builds topic clusters and 90-day editorial calendars against your actual brand system — set up once, consistent across every writer and every chat.
View Lena — Content Marketing Manager →Doing the math on your own workflow
You don't need exact pricing to know whether this applies to you. Ask three questions instead:
- How many fresh AI chats do I start in a typical week for repeat tasks — writing, analysis, support replies, anything with a "here's how we do this" preamble?
- How long does it actually take me to re-establish context each time, including the time spent finding the right doc?
- How often does the output miss the mark because the context I gave it that day was incomplete or slightly different from last time?
If the honest answers are "more than a few," "more than a minute," and "more often than I'd like," the fix isn't a better prompt. It's not writing that prompt again at all.
In summary:
Re-explaining your brand voice and role at the start of every AI chat quietly costs you tokens, time, and — most importantly — consistent output. A skill file like Jake — Conversion Copywriter or Lena — Content Marketing Manager is written once and reused in every session, so the AI shows up already knowing who it's supposed to be.
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 →FAQ
How many tokens does re-explaining brand voice actually cost per chat?
It varies with how detailed your instructions are, but a typical tone-plus-audience-plus-format context dump lands somewhere around 300-600 tokens once you include the natural preamble around it. That's the overhead you pay before you even ask your real question — and you pay it again in the next chat, and the one after that.
Isn't it cheaper to just keep the instructions in a saved doc and paste them each time?
It's better than retyping from memory, but it still isn't free. You're still paying the same tokens on every chat, still spending time finding and pasting the doc, and you still risk pasting an outdated version or forgetting a section. A skill file removes the paste step entirely — the instructions are already loaded before the chat starts.
Does this only matter for heavy AI users, or does it add up for occasional use too?
It scales with frequency, but the consistency cost shows up even at low volume. If you only start a fresh chat once a week to write copy, the token and time cost is modest — but the output can still drift depending on how completely you remembered your own brand rules that day. The more people on a team doing this independently, the faster brand consistency erodes, regardless of how "occasional" any one person's use is.