How to run daily local puzzles alone: generate, check, then approve

Filtering questions has become more important than creating them

Daily Who, made by an independent developer in Japan, presents one historical figure each day at midnight. Players can ask up to five questions in Japanese. Rather than creating each day’s question at the moment it is published, the service prepares 30 days of questions in advance.

The service does not stop at telling AI not to reveal a person’s name. After an answer is generated, it runs another check for the person’s name and aliases, then blocks answers that contain them. It places a separate safeguard between question generation and publication.

The daily word puzzle Purdle works similarly. When AI proposes word candidates, the service checks whether they overlap with previous questions, are too uncommon, or contain profanity. Only candidates that pass are sent to players.

A study of word-group puzzles compared 20 human-made puzzles with AI-made puzzles created through several methods. In roughly half of the comparisons, AI puzzles were rated as equally or more fun, creative, and difficult than human puzzles. But the production process also included generator and editor roles, along with human evaluation.

Endless Trivia has launched.

Narrow it to café training in Busan

Consider a manager training 12 café staff members in Japanese customer service in Jung-gu, Busan. They want to send one short daily question based on situations that repeat in the café: explaining menu items, checking allergies, or giving directions.

Today, the manager asks for Japanese sentences, checks a dictionary again, writes down Korean meanings, and creates three answer choices. They then check whether two answers could both be read as correct and post the question in a group chat. Even one question requires moving the same information across several screens.

If you build a service, the first thing to create is not an AI button. It is a “question library.” Keep the date, café scenario, Japanese sentence, accepted answer, source for the meaning, and publication status in one place, and build up 30 days of questions in advance. If today’s question fails, the service should be able to send an already approved backup question.

AI should only add drafts to this library. The system should check character count, Japanese writing format, duplication with recent questions, whether the answer is exposed directly in the hint, and whether there is only one accepted answer. Sentences where mistakes could create serious problems, such as allergy or safety guidance, should be excluded from automatic approval.

Instead of a long scorecard, the review screen should show the reason for failure immediately. The required action should be clear: “The same expression appeared within the past 14 days,” “Two choices both sound natural,” or “The explanation differs from the dictionary meaning selected as evidence.” Only questions that pass should be solved again in a phone-sized preview.

Regenerate only the failed question, not the whole batch. At the start, you can set a generation limit such as 10 attempts per day. If the same question fails twice in a row, stop generating and send it to a list for human revision. Ten attempts is a suggested starting point; the actual limit should be adjusted after recording per-question usage costs and failure counts for one month.

Just before publication, the manager still needs to read and solve the question. A sentence can be grammatically correct but sound rude to an actual customer. A word that is easy for beginners can also have a different meaning in a particular trade. AI reduces preparation time; it does not replace the person who approves language used in the workplace.

When you narrow the scope to cafés in Jung-gu, Busan, you do not need to solve the same problem as a nationwide Japanese-learning app. The first question set can fit actual menus and common situations, and you can meet the first users in a single visit. As staff members’ incorrect expressions accumulate, they can become material for next week’s questions.

Services that have tried this first

New Zealand’s Decryptic creates daily cryptic crossword clues with AI. It uses verified grid templates and automated structural checks, but the developer reads every clue and solves each puzzle before publication.

It launched on Google Play after a private test, and the day’s puzzle is free. A one-time payment removes ads and unlocks past puzzles. The developer said the biggest bottleneck was separating clues that are technically correct from clues that are fun and fair.

India’s solo-made service Grid Genius creates crosswords with AI after a player selects a theme and difficulty. Its creator says they continually adjust the generation instructions and algorithm, then solve improved puzzles themselves before distribution.

It offers free daily puzzles and currently operates on the web and in iPhone and Android apps. It offers subscriptions or separate access passes for unlimited puzzles, extra hints, and customization items, but has not disclosed user numbers or revenue.

The US local-place guessing game Pinpoint used AI only while preparing its place list, not during play. It did not use AI output to determine correct answers.

According to the creator’s account, AI usage for preparing one place cost less than one US cent. Place information was stored and reused. Controllable costs and attracting a large audience are separate problems.

Things you can build from this now

A three-minute Japanese question for cafés in Jung-gu, Busan

This is a staff-training service that sends one daily question about menu ordering and giving directions. It is for café owners and part-time staff who are beginning to serve Japanese customers.

Combining AI drafts with notation and duplication checks can reduce the burden on an owner who would otherwise write a new question every day. The first screen includes today’s customer scenario, one Japanese sentence, three response choices, and an owner approval marker.

Today’s quiz built from local news

This service creates one daily question from the original text of local news outlets or district office updates, then shows the supporting sentence after the answer. It is for district-level online newspapers where reporters do not have time to create a separate game every day.

Linking each question to its original article makes it easier to check whether AI added facts outside the article. The first screen includes today’s question, four choices, and a button that opens the original article after the answer is revealed.

A pre-shift safety question for a parts factory in Ulsan

This service creates one question before each shift from a factory’s safety rules and work instructions, then distributes it after approval. It is for safety staff and floor supervisors at automotive parts factories who repeatedly explain paper training materials.

Using only internal documents as evidence, and showing the paragraph that contains the answer, can reduce the risk of mixing in irrelevant general knowledge. The first screen includes today’s task, an illustration or sentence describing a hazardous situation, action choices, and a button to view the supporting guidance.

A daily walking challenge in Wolmyeong-dong, Gunsan

This service gives a short daily walking task using local places and things to observe. It is for organizers and participants in Wolmyeong-dong walking groups who have grown tired of the same routes.

Checking place coordinates and public access hours first, then changing only failed tasks, can help manage a small inventory of challenges even in a small area. The first screen includes today’s search area, the object to observe, a map, and a button to record completion.

What to ask three cafés today

Call three cafés in Jung-gu, Busan. Ask for one Japanese customer-service situation that has repeated recently and how they currently teach staff to handle it. Then ask whether they would be willing to try a three-minute daily question for one week without signing up.

If two of the three cafés can describe a repeated situation specifically and agree to a trial, it is worth creating the first seven questions. If every situation differs or they do not see a training need, narrow the trade further than the location, or switch from Japanese to a training topic that truly repeats.

Why this matters where you are

You can check whether a specific workplace in your market has a repeated situation that can become one short daily question. The language, safety rules, and local source materials will differ, but the workflow remains testable: prepare a question library, check AI drafts, and require human approval before publishing. Start by asking a small number of real workplaces whether they would try it.