Most people I know use AI writing tools the same way: describe the situation, ask for a message, paste it into email or a chat window, hope for the best.
That works for birthday wishes. It fails badly the moment money and contracts are involved. I learned this the hard way while helping a friend fight a 4,800 CNY deduction from her rental deposit, and it reshaped how I build every prompt in PromptVault.
The failure mode: plausible text is not defensible text
Ask a model to "write a message to my landlord about the deposit" and you get something fluent, polite, and completely useless in a dispute. It has no date anchors, no clause references, no request for evidence. The other side reads it, shrugs, and stops replying.
The gap is not writing quality. The gap is that the model has nothing to be accountable to. So the first fix is not a better instruction, it is a better input: paste the actual contract clause, the actual dates, the actual list of deductions. Then the model has something to reason over instead of something to imitate.
I ended up with a rule I now apply to every practical writing task: never ask for a document, ask for a position plus its evidence.
Guide one: the rental deposit dispute, split into three stages
The rental deposit guide grew out of that one case. The insight was that "getting the deposit back" is not one conversation, it is three, and they have incompatible tones:
- The notice. Establish that you gave notice within the contractual window, in writing. Facts only, no grievance.
- The inspection objection. Do not argue that wear is normal. Force itemisation: which wall, how many square metres, what unit price, which contractor, is there a quote. Ask for evidence, not fairness.
- The formal demand. Switch channels from chat to a letter with a deadline, sent both by message with a screenshot and by registered mail to the address on the contract.
Each stage has a copy-ready prompt that specifies the role, the allowed inputs, the output structure, and the prohibitions. The prohibitions matter more than the structure. In the rental guide the prompt explicitly forbids citing any statute the user has not supplied, forbids inventing amounts, and forbids emotional language, because a single fabricated article number hands the other side a reason to dismiss everything else you said.
The full three-stage workflow, including the inspection checklist and the clause comparison table, is here: https://promptvault-c25.pages.dev/guide/rental-deposit-dispute-guide
Guide two: the performance self-review
The second guide applies the same logic to a completely different anxiety: writing a self-review that survives a calibration meeting.
Most self-reviews fail because they describe effort instead of evidence. "Led the migration project and collaborated across teams" is a feeling. A reviewer cannot defend it in a room full of people comparing numbers.
The fix is an evidence chain: claim, measurable change, baseline, and the artefact that proves it. The prompt asks the model to interrogate the user rather than write for them, which is a subtle but important inversion. It refuses to produce a sentence until it has a number attached to it.
That guide, with the question set and three worked examples, is here: https://promptvault-c25.pages.dev/guide/performance-self-review-guide
What transfers across both cases
Three instructions show up in almost every guide I write now.
Constrain the sources. Tell the model which documents it may use and forbid everything else. This is the single highest-value sentence in any professional prompt, because hallucinated authority is the most expensive kind of error.
Put the blanks in the prompt, not in the model's imagination. Ask for placeholders like [INSERT DATE] instead of letting the model fill a plausible value. If a number must appear in the final text, a human should type it.
Ask for the next action, not the conclusion. A good output ends with a concrete request: three working days for a quote, a joint inspection, a written itemisation. Conclusions end conversations; requests start them.
A note on tone
Tone is not decoration. In the deposit case we moved deliberately from warm to formal across the three stages, and the shift itself did work: the landlord answered the second message in two days and the third within one. Formality signals that the conversation now has a paper trail. Warmth in the first message keeps the door open for a cheap settlement. Deciding which of the two you need, stage by stage, is worth more than any single cleverly worded sentence. Most guides skip this and hand you one "professional tone" for everything, which is why they read like they were written by a lawyer who has never had a landlord.
Where this is going
PromptVault is a free collection of these practical guides. No prompt packs, no course funnel, just the workflows I needed and could not find written down properly. The index is at https://promptvault-c25.pages.dev/guide , and new guides go up roughly once a week.
If you take one thing from this post: the next time you open a chat window to ask AI for help with something that involves money, a contract, or a performance rating, paste the source document first. The quality of the answer is set by the quality of what you hand over, not by how politely you ask.
Top comments (0)