Notes from the bench.
This is the thinkhard.si blog, on writing, testing and paying for prompts at work. It is for managers, marketers, support leads, founders and developers who use AI in real jobs. Each post takes a position, shows its arithmetic and links its sources.
Nine posts so far. A new one every Tuesday.Subscribe with RSS
- AI for DCF valuation: what to delegate, what to checkA language model reads filings and argues with your story well, and it should never compute a discount factor or supply a share count.
- Why your AI writes like a brochure, and how to stop itBrochure prose fills the space where facts should be, so take the style words out of your prompt and put the facts, the reader and a length in.
- Chatbot system prompt: write it like a contractScope, sources of truth, promises the bot may not make, refusals, escalation and injection defenses, written as clauses you can test.
- What a support chatbot costs at 10,000 chats a dayTurn-by-turn arithmetic on six models shows why conversation history drives the bill, what caching and batch save, and why cost per resolved conversation is the number to watch.
- How to choose an LLM: the cheapest model that passesSet the bar first, test across price tiers and count what failures cost, because at list prices the same job can cost 100 times more on a top model.
- Prompts are the new standard operating proceduresThe prompts that answer customers and route tickets are company procedures, so give each one an owner, a version, a test suite and a review date.
- Why AI answers change every time, and when it mattersTemperature zero will not make answers repeatable, so decide where variation is a bug and test across many inputs and repeated runs.
- “Make it better” is the most expensive prompt you writeVague prompts cost hours of rework and confused readers, and a short brief that says what is wrong and what good looks like fixes both.
- Do you need an AI tool, or just a better prompt?Many AI tools are a hidden prompt and a text box, so here is how to tell when a good prompt is enough and when a product earns its price.