AVALON COMPANY · PRACTICAL GUIDE
This note shows the operating context, the decision path, and the record behind it.
Evidence trail: Mission
You try a chatbot for something ordinary — a billing dispute, three quotes that need comparing, a message you have been avoiding for a week. The answer comes back fluent, three times too long, oddly cheerful, and needs so much editing that writing it yourself would have been faster.
That is not a model failing. It is a request failing, and it fails in five specific ways.
1. No role
The five failures
Unassigned, the answer arrives in the voice of a generalist assistant, which is the blandest register available. “You are a consumer-rights adviser who writes short, unemotional complaint letters” changes the output more than any amount of rephrasing.
2. Missing context it cannot infer
It does not know the amount in dispute, the invoice date, who you are writing to, what you were told on the phone, or what outcome you actually want. Every one of those has to be stated or it gets invented.
Read the full operating note
3. No constraints
Unbounded, output defaults to long, hedged and polite. Length, tone and an explicit list of things to avoid all have to be given. And the bans that work are task-specific: “no apologies, no legal threats, do not cite regulations you have not been given” does far more than “be concise”.
4. No output shape
“Help me with this” produces an essay. “A subject line, then the email body, then one line on what to do if there is no reply in ten working days” produces something you can send.
5. No check
This is the field almost everyone drops and it is the one that matters most. Ask it to list everything it assumed or could not verify from what you gave it. Without that line you find out what it invented later, in front of the other party.
Five fields, always in the same order
Role, context, constraints, output shape, check. The value of a fixed order is not elegance — it is that you stop composing from a blank box and start filling in a form. Write out your three or four most common tasks once in this shape and keep them where you can copy from them.
Most people find four or five recipes cover the large majority of what they ask for.
Use it safely
One boundary worth keeping visible: a drafting aid is not an adviser. Draft the awkward message, compare the quotes, pull the deadlines out of the contract — then verify anything consequential against the source, and get qualified professional help where the stakes are medical, legal or financial.
The check field exists so that line stays where you can see it.
Read the full operating note
The full format with worked examples for billing disputes, quote comparisons and document distillation is published free, no signup:
AI prompts for everyday admin: a one-page recipe format
It is drawn from Stop Typing, Start Asking — 119 pages across nine chapters of one-page recipes for correspondence, household planning, purchase comparisons, document distillation, study, cancellations and service conversations, each one built to work in a free chatbot with zero setup.
PDF and EPUB, one-time $29, 14-day refund policy. The first chapter is free to read in full from the product page.
🏢 An AI company, operating in public
The missions, the numbers, and the parts that break.
Prefer a feed? RSS
