A 5-person team that lets everyone freelance their own AI prompts isn't just burning tokens — it's burning hours and shipping inconsistent work. This is a representative case study (a composite, not a real named client) showing what that waste actually looks like month over month, and what changes when the same team switches to a small shared library of AI skill files.
Illustrative case study · By the KissMySkills Team
The setup: a normal, ad hoc AI workflow
Picture a 5-person marketing and ops team at an early-stage startup — a content lead, a social media manager, a CRO/analytics person, an ops coordinator, and a founder who dips in and out of everything. Every one of them uses Claude, ChatGPT, or Gemini daily. Nobody set up anything formal. Everyone just... types.
That means every new chat starts from zero. Before the AI can produce a usable first draft, someone has to re-explain:
- Who the brand is and who it's talking to
- The tone rules ("more direct, less corporate," "never say 'delve'," "always end with a clear CTA")
- Formatting standards (headers, bullet style, word count ranges)
- Which channel this is for and what "good" looks like there
- Context about the current campaign, product, or quarter
None of that is written down as a reusable asset. It lives as tribal knowledge, half-remembered and re-typed — slightly differently — every single time.
The math: 5 people, multiple sessions a day, every day
These numbers are directional, not audited — the point is the shape of the waste, not the decimal precision. Say each person runs 4-6 AI chat sessions a day for work: drafting a caption, rewriting an email, checking a landing page headline, summarizing a call. If even half of those sessions start with a fresh "let me explain our brand voice and formatting rules again" block of 150-300 words, that's roughly:
- 5 people × ~3 context-heavy sessions/day = 15 redundant context re-explanations a day
- 15 × ~20 working days/month = around 300 redundant re-explanations a month
- At a conservative 200 words of re-typed context each time, that's on the order of 60,000+ words a month of pure re-explanation — context that produces zero new output, just re-establishes the starting line
Multiply that across input tokens processed on every single message in the thread (most chat tools re-send the running context on each turn), and the token overhead compounds fast. But the token bill is the smaller problem. The bigger one is time.
The real cost isn't tokens — it's minutes, multiplied by five
Say re-typing or copy-pasting a "here's our brand and formatting context" block takes 2-4 minutes per session, plus the extra back-and-forth when the AI's first draft misses the mark because the context was incomplete or worded differently than last time. Across a team of five, that's easily 1-2 hours a week per person spent re-establishing context that should have been settled once.
Over a month, that's a meaningful chunk of a full workday per person — gone not to doing the work, but to re-briefing the AI on how to do the work. On a 5-person team, that compounds to something close to a full extra person's worth of "context tax" every month.
Drop Maya into Claude, ChatGPT, or Gemini once and every caption, calendar, and platform-native post comes out in your voice — no re-explaining tone or formatting per session.
View Maya — Social Media Manager →The second cost: quality drift
Redundant tokens and lost time are the visible costs. The quieter one is drift. When five different people each write their own version of "here's our brand voice" from memory, they don't write it the same way twice — let alone the same way as each other. One person's ad hoc phrasing leans more formal, another's drops in slang the brand doesn't actually use, a third forgets to mention the CTA rule entirely.
The result: content, emails, and analysis all technically "on brand" but each carrying a slightly different accent. Customers and prospects notice this even when nobody on the team can quite articulate why something feels inconsistent. It's death by a thousand small deviations.
This shows up hardest in two places: the CRO/analytics workflow, where inconsistent framing of "what's a meaningful lift" leads to different people drawing different conclusions from the same data, and marketing automation, where sequence logic and lead-scoring assumptions get reinvented informally every time someone new touches a campaign.
Beck diagnoses before he tests — funnel drop-off analysis, A/B test design, and result interpretation that stays consistent no matter who on the team is running the session.
View Beck — CRO Specialist →What changes with a shared skill library
The fix isn't a new tool or a new process doc nobody reads. It's replacing ad hoc prompting with a small, shared library of skill files — one per function the team actually uses AI for. A skill file is just a saved system prompt: brand voice, formatting rules, role context, and house standards, written once and pasted in at the start of any Claude, ChatGPT, or Gemini session (or saved as a Claude Project instruction so it's always loaded).
For a 5-person marketing/ops team, that typically looks like 3-4 skill files covering the recurring jobs: content and social, CRO/analytics, and marketing automation/lifecycle. Everyone on the team uses the same file for the same job instead of writing their own version each time.
What actually changes:
- Redundant context re-explanation drops to near zero — the brand voice and formatting rules are already baked into the skill file, not re-typed per session
- Output consistency goes up — five people running the same skill file produce work that sounds like it came from one team, not five individuals
- Onboarding gets dramatically faster — a new hire gets handed the skill files instead of a scavenger hunt through old Slack threads trying to reconstruct "how we usually write these"
- Time per session shrinks — less setup, less re-drafting because the first output missed unstated context, more usable output on the first pass
Finn designs systems, not one-off sequences — workflow architecture and lead-scoring models the whole team can reuse instead of rebuilding logic from memory each time.
View Finn — Marketing Automation Specialist →Onboarding: the underrated win
The token and time savings matter, but the onboarding effect is what teams notice most once they've made the switch. Instead of a new hire spending their first two weeks absorbing "how we do things here" through osmosis and getting corrected on tone in every review cycle, they open the relevant skill file on day one and their first output is already close to on-brand. The institutional knowledge that used to live only in senior people's heads — and leaked out of the org whenever someone left — is now a portable file that travels with the team.
If your team touches more than one of these functions, it's worth browsing the full Marketing & Ads skill library rather than building one-off prompts function by function.
In summary:
A 5-person team re-explaining brand voice and formatting from scratch in every AI session is quietly losing hours a week to redundant context and shipping inconsistent output as a result. Standardizing on a small shared library — Maya for content and social, Beck for CRO and analytics, and Finn for marketing automation — cuts the re-typing, tightens consistency, and turns onboarding into handing a new hire a file instead of a Slack archaeology project.
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
Is this based on a real client's numbers?
No — this is an illustrative, composite case study built from typical patterns we see in small marketing/ops teams using AI chat tools ad hoc, not a real named client's audited data. The token and time figures are directional to show the shape of the waste, not exact measurements.
How is a skill file different from just writing a good prompt once and reusing it?
It's the same idea, formalized and shared. The problem most teams have isn't that nobody ever writes a good prompt — it's that each person writes their own version, it lives in their personal notes, and nobody else on the team uses it. A skill file is a single, maintained version the whole team pastes in or loads as a Project instruction, so everyone starts from the same baseline instead of five slightly different ones.
Do skill files work outside of Claude?
Yes. Skill files are markdown-based system prompts, so they work anywhere you can paste in instructions — Claude Projects, ChatGPT custom instructions or Custom GPTs, and Gemini Gems. The team doesn't need to standardize on one AI tool to standardize on one shared context.