[Beginner] — assumes note 02. This is “prompt engineering,” minus the mystique. You have a text box, not an API — everything here works in one.
The kid version first
Imagine briefing a talented freelancer who starts work the instant you stop talking, never asks a clarifying question, and cannot see your screen, your company, or your last email.
They’ll do something regardless of how vague you were. The quality of what you get back is mostly the quality of the brief.
That’s it. Prompting isn’t magic words. It’s briefing.
The four things a good brief contains
Most weak prompts are missing one of these. Most strong prompts are just an ordinary sentence with all four.
| Element | The question it answers | Example |
|---|---|---|
| Task | What do you want done? | ”Rewrite this email” |
| Context | What do you need to know first? | ”It’s to a client who missed two deadlines; I want to keep the relationship” |
| Format | What should the output look like? | ”Under 150 words, no bullet points, warm but direct” |
| Audience | Who reads it? | ”She’s technical, senior, and dislikes padding” |
Compare:
Make this email better.
Rewrite this email to a client who’s missed two deadlines. I want to be direct about the impact without damaging the relationship — we want to keep working with them. Under 150 words, plain prose, no bullet points. She’s senior and dislikes padding.
Same effort, roughly. Wildly different output. Nothing clever happened — the second one simply removed the guessing.
The techniques worth knowing (there are only about six)
Show an example. Describing a format abstractly is harder than demonstrating it. Paste one instance of the output you want — one summarized meeting note, one rewritten headline — and say “like this.” Pattern-matching is what the machine is built for, so this is disproportionately effective.
Ask for reasoning on hard things. “Work through it step by step before giving the answer.” As note 02 explains, its own words are its scratch paper. Skip this on simple tasks — it just makes them longer.
Give it a role, lightly. “You’re an experienced GP explaining this to a worried patient” shifts vocabulary and framing usefully. Don’t overinvest: it changes tone and register, it does not install expertise. A role prompt cannot make a model know something it doesn’t.
Ask it to interview you. Genuinely underused and the best single tip for a beginner:
Before you write anything, ask me the five questions you most need answered to do this well.
It flips the dynamic from “guess what I want” to a briefing conversation. Particularly good for writing tasks, planning, and anything personal to your situation.
Iterate instead of restarting. Your first prompt is a first draft. Say what’s wrong — “too formal,” “you invented a detail in paragraph two, drop it,” “keep the structure but halve the length.” Steering beats rewriting the whole prompt from scratch.
Ask for options. “Give me three versions with different tones” costs you one message and gives you something to react to. Reacting is much easier than specifying.
What doesn’t work
- Politeness as leverage. Please and thank you are fine — they’re not doing any work. (Being specific is doing all of it.)
- Threats, urgency, fake stakes. “This is critical for my career” is folklore. Skip it.
- Insisting harder. If it can’t do something — current events without search, reliable arithmetic, knowledge of your private files — no phrasing fixes it. That’s a wrong-tool problem (note 03).
- “Are you sure?” as a check. It tests nothing; it just makes a concession more likely. To really test an answer, ask it to redo the work, or ask again in a fresh chat.
- One giant prompt with fifteen requirements. Long stuffed instructions reliably drop several. Break it into steps.
Failure modes → the actual fix
| What happened | Why | Fix |
|---|---|---|
| Bland, generic output | Nothing to condition on | Add context and audience; give an example |
| Ignored half your instructions | Too many at once | Split into two messages |
| Confidently wrong facts | It’s guessing from memory | Give it the source, or use a searching tool |
| Wrong tone | You didn’t specify one | Name the tone, or paste something in the voice you want |
| Weirdly formal, lots of “Certainly!” | Default assistant register | ”Drop the preamble, answer directly” |
| It agreed with something you know is wrong | Agreement is the likeliest next text | Ask it to argue the opposite case |
That last row is worth internalising. These systems lean agreeable. If you want a real critique, you have to ask for one explicitly — “what’s the strongest objection to this?” — because “this is great” is always available as plausible text.
Key insight
Every prompting technique is the same technique: reduce what the model has to guess. Context, examples, format, audience, and reasoning room are five ways of doing that. If output is bad, ask “what did I make it guess?” before reaching for a cleverer phrasing.
Related
- How LLMs Work, Plainly — why removing guesswork is the whole game
- Context and Long Chats — the other half of a good brief: what’s in the window
- Verifying What It Tells You — a well-briefed answer still isn’t a checked answer
- Prompt Engineering — the builder’s version: system prompts, chaining, evals, optimization
Software engineering
Infrastructure & security
Computer science
Hardware & engineering