Text is the basic discipline
Emails, posts, descriptions, reports, translations — the most common use of AI and the best practice ground for the prompting from the previous module. This is also where AI gives the fastest win: routine text that used to take half an hour shrinks to five minutes.
The key mindset shift: the model is an editor and co-author, not a "push a button, get output" machine. The best results come from your material (facts, a draft, a voice dictation) run through the model.
Emails and messages
The working scheme is to give facts and a frame, not to ask "write an email":
Write a reply to a client. The situation: they're asking for a 20% discount, we can give 10% at most with a year's prepayment. The client is important, the relationship is good, we can't refuse point-blank. Tone — warm but not fawning. Length — up to 120 words.
A quick trick for unpleasant emails: dictate or write everything you think, uncensored — and ask "translate this into a professional tone, keep the substance." Emotion turns into arguments.
Editing: the strongest mode
The model as editor is more reliable than the model as author: your draft contains your thoughts and facts, the model improves the form — there's almost nowhere to hallucinate here.
- "Cut it in half without losing meaning" — the best test of a text's quality: it instantly shows where the filler is.
- "Find the weak spots in the argument, be picky" — a review before sending something important.
- "Rewrite for audience X" — the same meaning for a director, a client, or a teenager.
- "Check spelling and punctuation, list the edits separately" — control over what exactly changed.
A voice draft is the fastest pipeline for long texts: dictate a stream of thoughts into the mobile app (5 minutes), ask it to structure and tidy them up. Typing structured text from scratch takes many times longer.
Posts and public texts
The problem with generation "from nothing": you get a smooth, faceless text that reads like a detergent ad. Readers have learned to spot AI texts with no personality instantly.
The cure is your own material as input: a specific case, numbers, a personal opinion, a detail only you know. The model will package the material into a form, but the material has to be yours. Plus examples of your old posts to convey the style (the few-shot lesson).
A good division of labor: headlines and opening variants — for the model (generating 10 options and choosing is its strong side), the final voice of the text — yours.
Translation
LLMs translate by meaning, not by words — head and shoulders above classic translators on living texts: idioms, tone, cultural references. Techniques:
- Set the register: "translate for business correspondence" and "translate colloquially" give different texts.
- Localization instead of translation: "adapt for an English-speaking audience, replace metaphors with ones a native speaker will get."
- A back-check: translate the result back with another chat and compare the meaning, if you don't know the language.
Key takeaways
- Give facts and a frame — the model builds the text from your material, not from emptiness.
- Editor mode is more reliable than author mode: your draft + the model's edits.
- A public text without your own material is faceless noise; material as input is mandatory.
- Voice dictation + structuring is the fastest path from thoughts to text.