Trust · how these are checked

What “checked” means here.

Not “AI-powered quality assurance”. The library is written and read by a person, and a set of mechanical rules runs on every build and refuses to publish a prompt that breaks them. This page says exactly what is checked, what is not, and how to test one prompt yourself in ten minutes.

Build checks, every build No invented test scores Last updated 24 October 2026 What still goes wrong →

The build stops if a prompt breaks a rule

These run on every build of the site, over all 3,200 prompts. If any one fails, the build writes nothing and the site is not updated. They are crude checks — a machine cannot tell whether a prompt is any good — but they catch the mistakes that are easy to make when you write a thousand of them.

A unique title. Two prompts with the same title means one of them is a duplicate nobody noticed.
At least 100 characters of instruction. A one-line prompt is a wish, not a brief.
The three metadata fields. When to use it, what it is tagged with, and the level. A prompt with no “when” cannot be found by the job it does.
At least one [PLACEHOLDER]. The prompt has to show the facts it needs, in capitals, where they go.
An ask-don’t-invent instruction. The body must actually contain wording that tells the assistant to ask, confirm or list assumptions. It is a text match, not a judgement, but a prompt without it never gets published.
The counts. No trade over 100 prompts, exactly 32 trades, exactly three free samples with full text. If the free tier or the trade count drifts, the build fails rather than quietly changing the offer.

Kits carry the same shape — name, when to run it, the prompt steps in order, a worked input and output, the checks and the limits — and are read by a person against that shape. They are not machine-scored, and nothing on this site claims they are.

What a person checks

Four questions, in this order, for every prompt and kit:

Is it a real job?

Not a topic, a task. “Chase an unpaid invoice without burning the customer” is a job. “Business writing” is not. If nobody would search for it on a bad Tuesday, it gets rewritten or cut.

Does it name the facts it needs?

The placeholders have to be the things only the owner knows: the real price, the real slot, the terms, the name of the scheme. If the prompt could be answered without them, it is too generic.

Does it say where the judgment stays?

Prices, refunds, hiring, discipline, health and legal wording are decisions, not drafts. The prompt has to say so, in the prompt itself, not only in the marketing around it.

Then the fourth: whether the example output is honest. A worked example that quietly invents a price to look complete is worse than no example.

What is not tested

No model-vs-model benchmarks. The prompts are written to work in the free tier of the common assistants. There is no score comparing them, and a score would be meaningless across models that change monthly.
No accuracy figure. Nobody can honestly publish “94% accurate” for a writing prompt. The output is a draft; the check is a person reading it.
No measured time savings. The site publishes none, and this page will not invent one.
No customer results. The worked examples are written examples, labelled as written examples, not transcripts of a real session or a customer's outcome.
No guarantee the output is right. The whole design assumes it might not be. That is what the brackets and the checks are for.

Test one yourself, in ten minutes

This is the test the free prompts are there for. You do not need an account, and you do not need to trust this page.

Pick a job you did this week — a reply, a quote, an advert — and write down the facts: the price, the time, the name, the promise. Keep that list.
Run the free prompt with the facts you are willing to share, with names and addresses taken out. The three full samples are here.
Compare the draft with your list. Anything in the draft that is not on your list was invented. That is the only accuracy measure that matters for your business.
Check the brackets. A good prompt leaves a visible gap where it was not told something. A prompt that fills every gap is guessing.
Run the draft through the free checker at /tools/check with your facts pasted in, and see whether it catches what you caught.
Decide. If the draft saved you a real edit, it is useful. If you rewrote it from scratch, it is not — for that job, in your trade, at your prices.

When a prompt fails

It goes on the failure log, with the guard that should have prevented it and the check that catches it. That log is the honest record: nine failure modes so far, each one written into the prompts rather than smoothed over in the marketing. If a draft fails in a way that is not on the list, send the shape of it, with the customer detail removed.