$3 a Month Bought Our Team a Prompt Style Guide

Most teams never write a prompt style guide because it sounds like paperwork that pays back never. It pays back in tokens. On a pay-as-you-go relay, every vague prompt is a longer answer, every longer answer is a bigger line on the receipt, and every instruction you forget to save is re-typed and re-sent on the next call. CheapRelay bills per real token, including cache reads and cache writes, with no monthly fee and no subscription, so the cost of a sloppy prompt shows up the same day you write it.

This walkthrough assumes three to six people, one shared API key, and a budget of about $3 for the month you spend writing and testing the guide. That $3 is a planning target, not a listed price. If your usage comes in under it, you keep the difference. If a rewritten prompt doubles the output, the number moves and you see exactly why. That feedback loop is the whole reason to write the guide.

01 / Lock a $3 monthly budget before writing the first rule

Decide the number first and let the number decide the rules. Three dollars per month is not much room for guesswork, so write down what it buys: one dollar for drafting and quick tests, one dollar for side-by-side comparison runs, and one dollar held back for the one deliberately messy prompt you run to see what the tidy version actually saves.

Keep a plain text file next to the guide with the date, the prompt name, the model, the input tokens, the output tokens and the cost of each run. Copy real numbers from your dashboard instead of estimating from memory. After two weeks the file will show which prompt is the expensive one, and it is almost never the one people complain about. Teams are usually surprised that the polite filler sentence at the top of every prompt is worth more than the whole feature description.

02 / Step one: collect the five prompts your team already sends

Do not invent rules for prompts nobody uses. Open the last week of chats, tickets and code comments and pull out the five prompts that appear most often. On a small product team those tend to be: write a customer reply, summarize a thread, draft a function, explain a stack trace, and turn rough notes into a changelog entry. Two people will have written near-identical versions of each one, which is exactly the duplication you want to remove.

Put the duplicates side by side and circle what differs. One says be concise, the other says nothing about length. One names the audience, the other assumes it. One pastes eight hundred words of context the model never needed. Now count the words. A 120-word prompt sent forty times a week is 4,800 input words before the model answers anything. Trimming it to 70 words removes about 40 percent of that input, and the saving repeats every week without anyone doing extra work.

03 / A four-part template anyone can fill in in two minutes

Every prompt in the guide should carry the same four parts in the same order, so a new teammate can copy the shape without reading a manual.

• ■ Role: one line stating who the model is acting as and for whom. Support agent writing to a non-technical customer is enough. Skip adjectives.
• ■ Task: exactly one verb. Reply, summarize, draft, explain. If a prompt needs two verbs, split it in two and let the person chain the results.
• ■ Context: only what the model cannot know. Product facts, dates, limits, the one constraint that changes the answer. Delete everything else.
• ■ Output: the exact shape. Word limit, sections, language, tone. Under 120 words is a rule you can check. Be concise is not.

Add one final line to every prompt: if the context is missing, say so instead of inventing it. That line costs roughly a dozen tokens at per-token billing, which is not free, and it is still cheaper than a fabricated date reaching a customer. Put it in the template so nobody has to remember it.

04 / Test every rule twice before it enters the guide

A rule that sounds tidy can cost tokens or damage the answer, so test before you standardise. Run the old prompt and the new prompt on the same input, twice each, and record input tokens, output tokens and the total. Keep a rule only when the answer is at least as good and the total drops. If the tidy prompt is worse, the rule is wrong or you cut context too hard. Add the missing line back and count again.

Two models are plenty for this comparison work. A small fast model such as Claude Haiku 4.5 or DeepSeek V4 Flash handles the bulk of the runs cheaply, and you only escalate to a larger model when the test shows the small one is failing a task. Save one good result and one bad result per rule inside the guide. Examples teach faster than adjectives, and they cost nothing to store.

05 / The weekly ten-minute review that keeps the guide honest

Every week, open the receipt file and ask three questions. Which prompt produced the largest total? Which rule did nobody follow? Which rule changed nothing in the answers but added tokens to every prompt? Delete that last category outright. A style guide grows more useful as it shrinks, and a one-page version that people actually open beats a ten-page version that sits unread.

Track three plain numbers per week: total tokens spent, tokens per prompt, and the share of prompts that follow the template. If the share drops below about 80 percent, the template is too long or the tooling makes it awkward. Simplify before you add another rule.

06 / One shared key across Claude Code, Codex and Gemini CLI

Keep the guide beside the code that uses it. CheapRelay gives one key that reaches a long list of models, works with Claude Code, Codex, Gemini CLI and any OpenAI SDK, and exposes OpenAI-compatible routes plus Gemini-native ones. That means the same prompt file can sit in your repository and be read by whatever tool the person on call happens to use, without rewriting the wording for each CLI.

• ■ Store the four-part template as a plain file in the repo so every tool can read it.
• ■ Store the tested examples next to it, with their token counts attached.
• ■ Keep the weekly receipt file in the same folder so cost is visible in the same place as the rules.

Start with five prompts, one page of rules and a $3 ceiling. If you want to run the tests yourself, sign up at CheapRelay and grab a key; the per-token billing means the first comparison runs cost only the tokens you actually use. Share the result with a teammate and you also get 20 percent of what invited users pay.