Case bank: before and after

Practicum section. Verified: 2026-06. These techniques come from communities where people share how they use AI. They are field self-reports, not verified business cases. The numbers are the participants' own estimates, and we did not verify them. Take the pattern, including how the person framed and checked the task, rather than treating the number as a promise.

The community uses one practical test: an ordinary person should be able to repeat the case from its description. You will not find elaborate setups below. Each case has four fields: input, process, result, and the point where a person must check the work.

Source: anonymized field reports from Telegram communities, February to June 2026. Every number below is a participant's estimate, not an independently confirmed result.

Work and business

1. Contract and procurement: build a position with legal sources

Input: contractor and customer terms and conditions, plus the country where the equipment was made and installed.
Process: AI compared the terms, found the applicable law, and drafted a position with quotes. A person edited and sent every reply.
Result (self-reported): a dispute that had lasted for months began to move. In a similar case, the annual contract price was revised downward. These claims come from the participants.
Review: check legal references and wording against primary sources or with a lawyer. You make the decision.

2. ICP by rejection: who should NOT become a client

Input: a prospective company's profile and your ideal customer criteria.
Process: instead of asking why the company was a fit, the person asked AI to find reasons to reject it, such as the wrong size, market, or product. This counters confirmation bias.
Result (self-reported): lead qualification became noticeably more accurate, according to the participant.
Review: weigh reasons for and against the client yourself. The model only adds possible reasons against.

3. Meeting preparation: rehearse the questions

Input: anonymized roles and materials for the participants, such as the CFO and operations director.
Process: the person gave the model those roles and asked it to play the other side and ask hard questions. They rehearsed the meeting, including by voice while traveling.
Result (self-reported): many of the real questions had come up in rehearsal. Time before the meeting became useful practice.
Review: this is a dry run, not a script. The real conversation will take a different path.

4. Extract numbers from a real message thread

Input: a message archive where amounts appeared in different formats and contexts.
Process: the model classified the amounts and put them in a table using one format.
Result (self-reported): one row out of about 30 was wrong in the sample that the participant checked.
Review: check a manual sample. Models confuse number formats.

Personal tasks

5. Claim for a defective product

Input: details about the product and defect, partly from connected email and cloud storage.
Process: AI collected the facts, outlined the steps, and prepared a claim draft. A person sent it.
Result (self-reported): a clean, nearly finished draft without collecting every fact by hand.
Review: verify the facts and recipient. You press the send button.

6. A purchase with tight limits

Input: the available space, compatibility with existing equipment, and budget.
Process: the person asked for an option that met every limit, not the "best" model. They also asked for one risk per option.
Result (self-reported): the choice worked. The limits mattered more than an abstract request to "find the best one."
Review: check price and stock with the store on the day you buy. Models invent both.

7. Personal memory: read-it-later that you can find again

Input: a stream of links that the person would usually save and lose.
Process: a bot saves each link, writes 10 to 20 lines on why it may help and what else it could be called, creates a daily digest, and supports semantic search.
Result (self-reported): it addresses forgetting rather than reading. Later, the person can find something they would not have remembered well enough to search for.
Review: the pattern has an obvious short-term use. The participant did not prove its long-term effect.

8. A knowledge base built from articles

Input: dozens of specialist articles and PDFs.
Process: the person put them in a local database, asked questions, and received reviews with formulas, ranges, and links to primary sources.
Result (self-reported): a review that used to take one or two days took about 30 minutes.
Review: the participant warns readers to check details and the underlying science. A model can summarize a source but is poor at extrapolation.

Failure cases: where the method breaks

These limits are not an argument against AI. They show where a prompt becomes useful only when a person remains responsible.

Use caution with	What field reports show
Medicine without a doctor	Even a user with medical training warns that a meaningful share of AI medical answers may be wrong. Use AI to prepare questions for a doctor, not to make a diagnosis.
Recording meetings without disclosure	Transcripts and summaries can help, but a hidden recording may be unethical or illegal. Check consent, NDAs, biometric data rules, and local storage.
Complex "autopilots"	As the setup becomes more complex and the goal moves toward removing people completely, failures and evaluation work increase.
Engineering specifications	AI-generated specifications can contain requirements that are physically impossible. Without an expert, the error may survive many review rounds.
Someone else's case is not your skill	A case library cannot replace the ability to use a tool, just as a phrasebook cannot replace a language. That is why this section teaches a method instead of "magic" prompts.

Send us your case

This bank grows through submitted techniques. If AI saved you hours, or failed with confidence, tell us about the task, prompt or technique, result, and how you checked it. A good case can be repeated by an ordinary person. A slogan without detail does not qualify, as the note at the top explains.

Contact the author on LinkedIn or use GitHub. We publish cases without names unless you ask us to include yours, and we mark each one as a participant's self-report.

Build your request with the templates and CLEAR method.
Analyze your role with the prompts for exploring your profession.

Version: 2026-07-09. Every number above is a participant's estimate, not an independently verified result. Check the facts, and do not treat someone else's metric as a promise about your outcome.