How to Write a Good Prompt: The Six Components
The short answer
A good prompt has six components: (1) role — whose perspective it writes from, (2) task — what you want, in one sentence, (3) context — for whom, in what situation, under what constraints, (4) examples — one to three real samples resembling the output you want, (5) format — length, structure, tone, which headings, (6) boundaries — what you don't want it to do. The two that make the most difference are examples and boundaries; most people write only the task and then blame the model. Practical rule: writing a prompt well once and turning it into a template beats rewriting it every time, on both speed and consistency. That's what "prompt engineering" amounts to in a business — not magic words, but reusable instructions.
Why do some people get better results from the same model?
Two people on the same subscription: one says "it does half my job", the other says "I tried it, it didn't work". The difference is almost never knowledge of the model; it's the ability to describe what you want. AI doesn't know the standard in your head — anything you don't specify, it fills in with the average. That's why you're getting average results.
I saw this most clearly while writing the How-To guides on this site, where I publish the exact prompts for the AI video effects I post on Instagram. At first I wrote short prompts and got a different result every attempt. The ones that ended up working all shared a structure: describe the scene, describe the motion, say what the camera does, say what must NOT change. That last item changed everything — without knowing what to hold constant, the model roams free.
The six components
Role — from whose perspective?
One sentence: "you're a sales manager with ten years' experience" or "you're an experienced editor". Its impact is overstated but not zero: the role sets which vocabulary and which priorities get chosen. Keep it short; a three-paragraph character sketch is wasted effort.
Task — what do you want, in one sentence?
Not "improve this email" but "rewrite this email so it chases an overdue payment without damaging the relationship". The task sentence needs a verb and a purpose. Without the purpose the model guesses your intent — usually wrong.
Context — for whom, in what situation?
Who's the recipient, what's the state of the relationship, what constraints exist. "A three-year client, first late payment, and we're discussing a new project next month" completely changes the tone of the output. Writing context feels slow but takes far less time than fixing the result.
Examples — the strongest component
Paste one to three past outputs you liked and say "match these". It beats ten paragraphs of description. The durable version in a business: keep a "good examples" file for every recurring task — quotes, emails, product copy. That file becomes your team's shared quality bar.
Format — what should the output look like?
Length ("150 words max"), structure ("three headings, two sentences each"), tone ("formal but not cold"). Without a format the model falls back to its default: medium length, bulleted, slightly too polite. The "AI smell" everyone complains about largely comes from that default.
Boundaries — what should it NOT do?
The most skipped and most useful component. "No emoji", "don't invent figures — leave blanks if you don't know", "don't slip into an apologetic tone", "no bullet points". The instruction "if you don't know, don't guess — write [MISSING]" in particular cuts error risk sharply in quotes and reports.
You don't need all six every time. For simple tasks, task + format is enough. But for recurring work where quality matters, write all six once and turn it into a template — write once, use a hundred times.
Bad prompt, good prompt
| Bad | Good |
|---|---|
| "Write me a proposal." | "Here are three of my past proposals; in that same tone, write one for social media management for a restaurant chain. 400 words max, three headings: scope, process, investment. Don't write a price — put [PRICE] there." |
| "Summarise this meeting." | "From this recording produce three sections: (1) decisions made, (2) who was assigned what and by when, (3) open questions. Add nothing that wasn't said; if unclear, write 'not clear'." |
| "Give me Instagram content ideas." | "My audience is 30–45, people running their own business. My three highest-engagement posts last month were [...]. Give me 10 Reels ideas in that vein; for each, describe what's on screen in the first 3 seconds. No generic motivational content." |
| "Fix this text." | "Simplify this: shorten sentences, cut adjectives, keep meaning and order intact. Preserve my sentences — don't rewrite, just remove excess." |
The prompts on the right look long, but each is written once and saved. On the second use you only swap the variables — client name, sector, numbers.
Four templates you can copy
Fill in the brackets with your own details. These are simplified versions of templates I use daily.
- PROPOSAL DRAFT — "You're an experienced sales lead in [sector]. In the tone of my three past proposals below, write one for [client type] covering [service]. Context: [what they want, what problem they have]. Format: max [X] words; headings: scope / process / investment / next step. Boundaries: don't invent a price — write [PRICE]; don't promise timelines; no inflated adjectives. Examples: [...]"
- MEETING NOTES — "From the recording below produce three sections: decisions / tasks (who, what, when) / open questions. Boundaries: add nothing that wasn't said, offer no interpretation, mark anything ambiguous as 'not clear'. Format: bullet list per section, one sentence per bullet."
- CLIENT REPLY — "You're not writing on my behalf, you're drafting. Context: [who the client is, relationship history, what happened]. Goal: [what you want to achieve]. Format: max 120 words, single paragraph, warm but professional. Boundaries: no apologetic tone, no commitments, no emoji."
- DOCUMENT ANALYSIS — "In the [contract/spec] below, find: obligations on us / penalty clauses / deadlines / clauses left ambiguous. For each finding, cite the clause it came from. Boundaries: no legal advice, nothing that isn't in the document, say 'ambiguous' when unsure."
At Gurizon we stopped keeping these templates in personal notebooks; we hold a brand-voice file per client and shared templates for recurring work. The reason isn't efficiency, it's consistency: when three different people write to the same client, it has to read like one voice. Moving prompts from personal to institutional is, in my view, the highest-return small change a team can make.
What to do when it isn't working
- Don't restart — correct. "This is good, but the second paragraph is too formal, redo it in this tone" gets there far faster than opening a new conversation.
- Put into words what you dislike. "I don't like it" isn't an instruction. "Too generic, no sector-specific detail" is.
- Give an example. If it's still off after three attempts, the missing piece is an example, not a better description. Paste something close to what you want.
- Split the job. Don't ask for research, writing and formatting in one prompt; break it into three steps. Quality drops most in long chains.
- Change the model. Try the same prompt in another assistant — for some kinds of work the difference is real. I compare which is good at what in a separate article.
Key takeaways
- Bad output usually isn't the model — it's an underwritten request.
- The two components that matter most: giving examples and stating what it should not do.
- "If you don't know, don't guess — write [MISSING]" sharply cuts error risk in quotes and reports.
- Write the prompt well once and templatise it; consistency is worth more than speed.
- When it misses, don't restart — correct it by naming what's wrong.
Frequently asked
Is prompt engineering worth learning, or will models make it obsolete?
The "magic words" side is already fading — models keep getting better at understanding poorly written requests. What doesn't fade is being able to describe precisely what you want. That's not a prompting skill, it's a thinking and writing skill, and no model can do it for you. What's worth learning isn't tricks; it's the discipline of writing down what you actually want.
Are long prompts better than short ones?
Length isn't a virtue; unnecessarily long prompts dilute attention. The rule: keep a sentence if it removes an ambiguity, cut it if it doesn't. Lines like "please be very careful and do your best" remove no ambiguity — delete them.
Should I write prompts in Turkish or English?
If the output is in Turkish, write in Turkish — removing the translation layer makes the text read more naturally. Exception: image and video tools. There, English prompts still produce noticeably more accurate results, which is why the prompts in my video guides on this site are in English.
How do I teach my team to write prompts?
Not by running a training — by handing over templates. Nobody applies anything after a two-hour presentation; give someone a working template for their own job and they use it the next day. What I do: write templates for the three most common tasks, put them somewhere shared, then ask everyone to adapt them to their own work. The learning happens in use.