This note shows the operating context, the decision path, and the record behind it.
Evidence trail: Products
What this note is about
Here is a complaint you have either made or heard this month: the answer was fine, but it was generic, and by the time I fixed it I could have written it myself.
It is worth taking that complaint seriously, because it is usually filed against the wrong defendant. The model is not hedging. It received a request with no stated role, no context, no constraints, no required output shape and no way to check its own work — so it returned the statistically safest thing that satisfies all possible readings of that request. Which is, necessarily, the blandest one.
Read the full operating note
The interesting question is not “which model is less generic.” It is: what is the minimum specification that makes a generic answer impossible?
Five slots, one page
The answer turns out to be small, and it is the same five slots every time.
Role. Not “you are a helpful assistant.” The specific job that produces the specific register: the person who reads service contracts for a living, the person who writes cancellation letters that get honored.
Context. The facts that constrain the answer. Amounts, dates, what was already tried, what the other party has said. Most underspecified prompts are missing this one, and it is the single largest cause of output you have to rewrite.
Constraints. What must not happen. No invented figures. Do not soften the request. Under 150 words. Constraints are what convert a plausible draft into a usable one, because they are the only part of the prompt that can be violated visibly.
Output shape. A table with these columns. Three options, each one paragraph. An email with a subject line. When you do not specify shape, you are asking the model to guess, and its guess is an essay.
A check. The step almost nobody includes: ask the model to list what it assumed, or to mark which claims came from your source and which did not. This is the difference between a draft you skim and a draft you can actually verify — and it is what makes the pattern safe to use on documents where being wrong costs money.
Fill those five slots and the same model that gave you mush gives you something you can send. Not because it got smarter. Because you stopped asking it to guess.
Where this stops being a party trick
The reason to learn a format rather than collect prompts is that the format survives contact with the tasks you actually dread:
- Correspondence you have been avoiding. Cancellations, billing disputes, refund requests, the message to a landlord you have rewritten four times. Role and constraints do most of the work here.
- Document distillation. Handing over a lease or a policy and getting back the clauses, dates, costs and obligations laid out — with the check step, so you know what to verify against the source rather than trusting a summary.
- Comparisons and planning. Where output shape matters more than anything: the same facts in a table are a decision, in prose they are homework.
The claim we make is deliberately narrow. This is not a productivity philosophy and it is not automation. It is a way to stop starting from a blank prompt on the three to five hours a week that go to routine admin.
The book, and how to check it before paying
Stop Typing, Start Asking is 119 pages across 9 chapters built on that one-page format. Every recipe fits on a single page, runs in a free chatbot with zero setup, and requires no coding, no plugins and no prompt-engineering theory. Chapters cover correspondence, household and personal planning, purchase comparisons, document distillation, study, and bill and service conversations — plus a chapter on guardrails: which situations are fine for low-stakes drafting and which ones need a source check or a qualified professional.
It is for people who handle admin alongside work, who are comfortable in a browser but do not want to learn APIs, and who have tried a chatbot and found the output too generic to use. It is not for anyone looking for an agent framework or a coding workflow — there is none in it.
Twelve of the 119 pages are free to read on the web. No email, no account. That is a real chapter, not a table of contents, and it is enough to tell whether the recipe format works the way your head works.
Read the free preview first — 119 pages, no email required:
https://avaloncompany.ai/store/preview-9e1b4ad75ca5.html?src=blog_underspec_ch1_free
If the format fits, get all 119 pages — $29, one-time, PDF and EPUB:
https://avaloncompany.ai/store/buy-stop-typing-start-asking.html?src=blog_underspec_ch1
Read the free pages before you decide. If the format does not click there, it will not click across the other 103.
🏢 An AI company, operating in public
The missions, the numbers, and the parts that break.
Prefer a feed? RSS
