Prompting, with evidence.
Practical guides for people who use AI at work: managers, support and sales teams, product managers and developers. Each one takes a single question, answers it with worked examples, and ends with the thinkhard.si tool that lets you check your own prompt.
- 1Write itClose the gaps that make a model guess.
- 2Lay it outShape the prompt for the model that reads it.
- 3Test itReal cases, exact checks, no regressions.
- 4Price itThe cheapest model that passes your tests.
- 5Harden itKeep secrets out and limit the damage.
New to AI
Start here if prompting is new to you. No technical background needed.
- Prompt engineering for beginners: the 15-minute versionThe five things every prompt needs, one office task rewritten before and after, how to fix a weak answer with feedback, and the habits worth dropping.
- How to write a prompt: brief AI like a new hireWrite prompts the way you would brief a capable new colleague: context, goal, audience, material, format, limits and what good looks like, with office examples.
- What is an API key? Why AI tools ask for yoursAn API key lets software use an AI model on your account. What it is, why tools ask for yours, what it costs in practice, free options and how to keep it safe.
- ChatGPT, Claude or Gemini for work: how to chooseWhat each of the three assistants offers as of October 21, 2026, what really decides the choice, and a one-hour bake-off on ten of your own tasks.
Using AI at work
Playbooks for people and teams who use AI every day.
- The ten checks of a well-written promptTask, purpose, material, scope, form, limits, standard, reference, doubt and fit: what each test asks, a weak and a fixed example, and how the score works.
- Choosing an AI model for business: cost, quality, speedPick an AI model for business tasks in plain language: cost, quality and speed, real prices per 1,000 tasks, and the cheapest model that passes your tests.
- AI prompts for customer support that follow policyA playbook for support teams using AI: replies that follow policy, triage, macros, tone and escalation, no invented refunds, and testing on real tickets first.
- Sales and marketing prompts that don’t sound like AIGet outreach and marketing copy from AI that sounds like you: feed it real facts, voice samples and banned phrases, ask for variants, and check every claim.
- Rolling out AI prompts to a team: a shared libraryTurn good prompts into shared templates a team can trust: an owner, version numbers, review, one library, tests before every change, and when to retire one.
- Using AI at work safely: a practical risk checklistWhat AI should and should not do in a business: confidential and personal data, invented facts, approvals, prompt injection, disclosure and human review.
Building AI features
For product managers and developers putting AI in front of customers.
- How to write a system prompt for a chatbot or assistantThe seven parts of a working system prompt, a support assistant rewritten before and after, and the tests to run before it meets a customer.
- Prompting Claude vs GPT vs Gemini: what differsWhat Anthropic, OpenAI and Google each recommend for prompt layout, what they share, and one prompt laid out three ways. Conventions checked October 10, 2026.
- How to test a prompt before you ship itTest a prompt without writing code: real test cases, exact checks, an AI judge you can trust, and a regression check each time the prompt or model changes.
- Extract data from invoices and PDFs with AIDefine every field, enforce the shape with a schema, let the model leave blanks, check the values in code and spot-check a sample, with a worked invoice.
- How to reduce AI hallucinations at workGive the model its sources, make “I don’t know” an acceptable answer, ask for quotes you can check, and test with questions the sources cannot answer.
- What an LLM call really costs: tokens and cachingHow tokens are billed, the formula for input, output and cached tokens, a worked example at current list prices, and how to pick the cheapest model that passes.
- Prompt injection explained: stop system prompt leaksWhat prompt injection and system prompt leakage are, why secrets never belong in a prompt, and the layered defenses that limit the damage when an attack works.
- How to measure an AI feature after launchFive numbers to track once an AI feature is live, how to collect each, what a bad trend means, and a weekly review that turns failures into test cases.
Plain claims, with sources
Prices come from the thinkhard.si model catalog and carry the date they were checked. What the guides say about Claude, GPT and Gemini comes from each vendor’s published guidance, linked where it is used. Example prompts and data are labeled as examples. If you find something wrong or out of date, tell us.