Prompt Engineering Basics
Seven principles that double answer quality
① Writing good prompts isn't about magic phrases — it's about reducing the blanks the model has to guess. More blanks, more generic answers.
② The seven: role, context, format, examples, constraints, verification, iteration. Among them, one example outperforms ten lines of description.
③ The most common mistake is using the first answer as the deliverable. A prompt is a conversation, not a command.
① Why the prompt changes the output
The model continues your input with "text that would plausibly follow". So your input is the specification for what kind of text should come next.
Type "tell me about stocks" and the model produces the most average continuation. On the internet, that question is usually followed by a generic primer — so you get a generic primer. The model isn't being lazy; the question specified that answer.
Writing a good prompt isn't "persuading the model" — it's filling in, up front, the blanks the model would otherwise fill by guessing. Fewer blanks means lower variance and an answer that lands closer to what you wanted.
② Principle 1 — Assign a role
The same question yields different content depending on who's answering. Naming a role narrows the style and knowledge domain the model draws on.
| Prompt | |
|---|---|
| Bad | Explain how to read financial statements |
| Good | You're a corporate credit analyst with ten years' experience. Explain to someone in their first job, reading a small manufacturer's statements for the first time, the order in which to look for signs of distress. |
One caveat: a role changes style and perspective; it doesn't conjure knowledge. Writing "you are the world's best analyst" does not improve data accuracy.
③ Principle 2 — Provide context
The model doesn't know your situation. Purpose, audience, what you already know, and your constraints will completely change where the answer aims.
| Prompt | |
|---|---|
| Bad | Recommend some ETFs |
| Good | I want to understand ETF types and structure. I have three years of single-stock experience and no ETF experience. Don't recommend products — explain why each selection metric (expense ratio, tracking error, volume) matters. |
Asking AI "what should I buy" is a structurally bad question. It doesn't know your assets, risk tolerance, or horizon, has no live prices, and above all is weak on numbers. Shift from asking for judgments to asking about structure and concepts and you land where AI is genuinely strong. For product-level comparison, use source-grounded material such as our ETF primer.
④ Principle 3 — Specify the output format
Without a format, the model defaults to explanatory prose. Tables, lists, numbered procedures, character limits — name the shape and you'll get it.
e.g. "Compare three items in a table. Columns: item / pros / cons / when to use. Keep each cell under 40 characters."
There's a side benefit: a table makes omissions visible. Prose can gloss over a gap; an empty cell cannot.
⑤ Principle 4 — Give examples (highest impact)
If you take only one of the seven, take this. Attaching one or two examples of the output you want communicates more precisely than ten lines of description.
| Prompt | |
|---|---|
| Bad | Make these headlines short and punchy |
| Good | Rewrite headlines in the style of these examples. Ex1: (before) Federal Reserve Decides to Cut Benchmark Rate by 0.25pp → (after) Fed Cuts Rates 0.25pp Ex2: (before) Nvidia Q2 Results Exceed Market Expectations → (after) Nvidia Beats on Q2 Now rewrite the following: … |
Why examples are so strong ties directly to how the model works. It's a machine for continuing patterns in its input, so showing the pattern is far more direct than describing it in words.
⑥ Principle 5 — Set constraints
State what not to do. Without constraints the model follows the average habits of its training data: long preambles, both-sides framing, and a closing "consult a professional."
Useful ones: "skip the intro, start with the substance", "three sentences per item", "nothing generic — only what applies to this case", "write 'unknown' if you're not sure", "include source URLs".
"Mark anything you're not certain about as uncertain."
This one line meaningfully reduces hallucination, because it explicitly gives the model permission to say it doesn't know. Without it, the model defaults to filling the gap.
⑦ Principle 6 — Make it verify
After the answer, follow up with "flag which parts of that answer need fact-checking and which claims have weak support." Models are surprisingly good at critiquing their own output. They can't do it while generating, but reviewing existing text is a separate task they handle well.
For load-bearing conclusions, layer on the five-step verification routine.
⑧ Principle 7 — Iterate
The most-ignored principle. Treat the first answer as a draft and point out specifically what's lacking.
"Try again" does nothing. Say "the third item is too generic — make it specific to 2026 conditions and leave the rest unchanged": name what, how, and what to keep.
1. Role — who is answering
2. Context — why, for whom, what you already know
3. Format — table / list / length
4. Examples — one or two target outputs (highest impact)
5. Constraints — what to avoid, mark uncertainty
6. Verification — have it review its own answer
7. Iteration — specify what / how / what to keep
⑨ Three common myths
"Longer prompts are better." No. Specificity matters, not length. Extra adjectives blur the target.
"Being polite gets better answers." Tone slightly affects the tone of the reply and has nothing to do with accuracy. Clear beats polite.
"There are magic phrases." Any given phrase's effect shifts with model and version. A structural approach like the seven above lasts far longer.
These seven principles are grammar that works across every task. What happens when you apply them to writing specifically is covered in Writing With AI — Prompts for Drafting, Editing, and Summarizing. The core point: drafting, editing, and summarizing need opposite amounts of latitude. For work you repeat often, reusing the seven principles instead of reassembling them each time is covered in building your own prompt templates.
※ This article covers general conversational-AI usage as of July 2026. Individual services differ in features. Investment-related mentions are illustrative only and are not investment advice.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.