Write Prompts That Work — From "Give Me Something Good" to "Large Jjajangmyeon, Hold the Chili"

OUTCOME — What you'll have afterward

Following this tutorial leaves you with three things.

What gets built

• One reusable prompt template — works for social posts, briefs, and document summaries alike
• A side-by-side comparison of the same task prompted badly vs. prompted well
• The habit of structuring any request into three blocks: context, instruction, format

📖 Term: What is a prompt? A "prompt" is any question or request you send an AI. The word comes from the theater — a prompter whispers missed lines from offstage. The mental model that clicks fastest is a restaurant order slip. Specific order → exact dish; vague order → whatever the kitchen felt like making.

Tutorial metadata

Why this matters — prompt quality = output quality

In the first tutorial I said AI isn't a perfect tool but a partner you learn to drive. About 80% of that driving skill lives in the prompt. With the same ChatGPT, a 7-character prompt and a 150-character prompt produce outputs that differ by 2x, 5x, even 10x in quality. That gap never closes with "a better tool." The answer only changes when the asker changes.

Where you are now: you know what this delivers. Next: prerequisites.

PREREQUISITES — What you need

Just three things.

Required

• A signed-in AI tool: ChatGPT, Claude Sonnet 4.6, or Gemini 3. If you don't have one, take 10 minutes on Pick Your First AI Tool first
• One real task of yours: not a made-up example — an actual thing you need to do this week
• A text editor or notes app: Notes, Notepad, or a Google Doc works

Not required

• Coding skill, English fluency, a paid subscription
• A "polished business plan" — fuzzy-state starting trains the muscle better

📖 Term: Does the free tier cover this? Yes. Every prompt fits inside any free tier's daily cap. Paid plans become worth considering when you spend more than two hours a day with an AI.

Environment check (30 sec)

Open your AI tool and confirm the input box is visible. If a prior conversation is loaded, click "New chat" — leftover context corrupts the "bad prompt comparison" in Step 1.

Where you are now: tool, task, and notepad ready. Next: why short prompts produce bland answers.

WHY — Why the prompt decides the outcome

The line I hear most from students: "I tried ChatGPT, it was just okay. Bland answers." When we open their screen the prompt almost always reads: "Tell me how to do marketing well." Seven words. The answer is generic — "Define your target. Use social media. Content marketing matters." Chapter 1 of any textbook.

Reason 1 — The AI isn't a mind reader.

Nobody walks into a restaurant and says "give me something good." You say "large jjajangmyeon, hold the chili." Somehow everyone forgets this the moment they face an AI. "Tell me how to market" is the exact equivalent of "give me something good." The AI doesn't know your industry, target, budget, or history, so it can only do one thing — return an answer that applies to everyone and therefore specifically helps no one.

Same AI, same moment. The only difference is one line of text.

The problem isn't the AI. It's the question.

Reason 2 — Constraints produce creativity.

This sounds backwards. People tell AI "write freely," thinking no limits means more originality. The opposite happens. With no constraints, AI takes the safest path — average structure, commonest phrasing — because not failing means not standing out. Constraints like "under 500 words, open with a question, include one statistic, end with a call-to-action" force the AI to solve a real problem inside boundaries. Character emerges from that demand. Same principle as 14-line sonnets and 5-7-5 haiku — the form doesn't cage the poet; it generates the tension.

📖 Term: What are "constraints"? In prompt vocabulary, "constraints" are rules about length, format, tone, what to include, what to avoid. "500 words," "skip the obvious" — all constraints. They aren't walls; they're filters that narrow choices and raise decision quality. Unlimited freedom converges toward average; well-designed constraints converge toward character.

Reason 3 — A vague prompt is evidence of vague thinking.

When you write "write something good," you haven't defined what "good" means to you. If the answer is "it depends on the situation," then describing that situation to the AI is literally what prompting is. Prompting is partly an act of looking at yourself: "What exactly do I want?" That question is the work.

Where you are now: three reasons the prompt decides the outcome. Now the 5-step build.

STEPS — Build a good prompt in 5 steps

Step 1. Send a bad prompt first (3 min)

You need the contrast to feel the force of a good one. In your open AI window, type this with a real task in mind:

Write me an Instagram post.

Hit Enter. Within seconds you get something like: "Hey everyone, today we're excited to share something special..." — text you could paste onto any business and it would half-fit. Don't close this answer. You'll need it for comparison in Step 5.

Step 2. Fill in the CONTEXT block (3 min)

Fill these four slots — one line per slot, off the top of your head.

[CONTEXT]
- Who I am: ___________________________
- Current situation: ____________________
- Constraints (budget, time, people): ___
- What I've already tried: ______________

A worked example:

[CONTEXT]
- Who I am: Solo founder making eco-friendly handmade soap
- Current situation: 500 Instagram followers, ~$400/mo revenue, only product photos
- Constraints: Near-zero ad budget, one-person team, weekends only
- What I've tried: 40 product photos, follower count flat

The point: the AI is a consultant you just met. The context block gets that consultant up to speed in 90 seconds.

📖 Term: "Context" — two meanings (1) In-prompt context: the background about who you are and your situation (this tutorial's focus). (2) Conversation context: prior messages ChatGPT / Claude / Gemini remember within a single chat. As of 2026, major AI context windows run from hundreds of thousands to over a million tokens — long conversations stay in memory from start to finish. Those first 5 lines fuel the next 50 turns.

Step 3. Fill in the INSTRUCTION block (3 min)

Mix in four ingredients:

• Length: "in 500 words," "three paragraphs," "five items"
• Structure: "problem → cause → solution," "three pros, two cons"
• Include: "specific numbers," "one real example," "a call-to-action"
• Exclude: "skip the obvious," "no greetings"

Continuing the soap example:

[INSTRUCTION]
Write an Instagram post announcing our new "Chamomile Calm" soap.
- Length: 4-5 lines of body
- Structure: question opener → product intro → call-to-action close
- Include: calming effect, natural ingredients — naturally woven in
- Exclude: ad-speak tone, hashtag dump

The point: see why "write something good" isn't an instruction? "Good" was never defined. These four slots make "good" measurable.

Step 4. Add FORMAT + ROLE + TONE (3 min)

Three more levers lift the result another step.

Format — same content, radically different usability.

Role — telling the AI "you are X" shifts perspective, vocabulary, and emphasis all at once.

You are a 10-year Instagram marketing expert.
You are a teacher explaining science to elementary students.

A role line replaces ten lines like "be professional · be practical · be detailed."

Tone — same info, completely different feel.

📖 Term: How role assignment works LLMs learned from trillions of documents spanning countless perspectives. Without a role, the AI averages them all, producing the blandest blend. Declaring "you are a 10-year marketing expert" narrows the mix to documents matching that role — expert vocabulary, practical framing, details a novice wouldn't catch follow automatically. Role is the lever of prompting. In 2026, models like Claude Sonnet 4.6, GPT-5.4, and Gemini 3 follow role instructions with high fidelity and maintain character across long outputs.

Step 5. Combine, send, compare (3 min)

Stitch the blocks into one prompt and send.

You are a 10-year Instagram marketing expert for small handmade brands.

[CONTEXT]
- Solo founder making eco-friendly handmade soap
- 500 followers, ~$400/mo revenue, only product photos posted
- Near-zero ad budget, solo, weekends only
- Follower growth flat, revenue flat

[INSTRUCTION]
Write an Instagram post for the new "Chamomile Calm" soap.
- Length: 4-5 body lines
- Structure: question opener → two benefits woven in → call-to-action
- Include: calming effect, natural ingredients
- Exclude: ad-speak, hashtag dump

[FORMAT]
Answer with body / hashtags (5 tags) / recommended post time — separated.
Warm but trustworthy tone.

Copy this, swap the [CONTEXT] lines for your situation, send it. Put the reply next to Step 1's output. The gap is almost uncomfortable.

Where you are now: one bad prompt, one good prompt, two outputs side by side.

VERIFICATION — What "done" looks like

All four should be "yes" before you call this complete.

1. [ ] The bad prompt's output is generic — could be pasted onto anyone's business. No specific brand, product, or target?
2. [ ] The good prompt's output is tailored to your situation. Are product name, target, and tone actually reflected?
3. [ ] You can explain the difference between the two in one sentence. If yes, you've internalized the core
4. [ ] The good prompt is saved in your notepad for reuse.

If any is "no," the common stalls: (1) [CONTEXT] was thin — go back to Step 2 and fill all four slots; (2) [INSTRUCTION] still contains vague words like "good" — go to Step 3 and make length, structure, include, exclude measurable.

You get what you put in. A 7-word prompt returns 7 words of quality.

VARIATIONS — When the situation differs

Variation 1 — Staged prompting

Writing a perfect one-shot prompt is hard. Split into three turns. Turn 1: context only, ask "Got it?" Turn 2: add instruction and format. Turn 3: react to the draft with "shorter here, more concrete there." The AI remembers prior turns in the same chat, so context doesn't need re-pasting.

Variation 2 — Build a template library

Save three to five prompts you use most often. I keep five: Instagram post, blog post, email reply, meeting summary, one-page brief. Each reuse is a 30-second edit of the [CONTEXT] slot.

Variation 3 — Change only the role

Keep the same context and instruction. Swap only the role line: "10-year marketing expert" → "fresh junior marketer" → "ruthless investor" → "friend in your target audience." Watching the same question produce four different answers teaches the power of role faster than any textbook.

Variation 4 — Long-document summary template

Read the document below and summarize in this format.
- One-line summary: (under 15 words)
- Three key points: (one line each)
- One actionable takeaway
- One missing piece or open question

[DOCUMENT]
(paste here)

Variation 5 — Structured JSON output

As of 2026, Claude Sonnet 4.6, GPT-5.4, and Gemini 3 all support reliable JSON schema output. Use this when you need to pipe results into a spreadsheet, app, or automation.

Return three Instagram post ideas as a JSON array.
Each item: { "concept": string, "body": string, "hashtags": string[], "postTime": string }

[CONTEXT]
(paste context block here)

TROUBLE — Common stuck points

❓ Writing long prompts feels tedious

Normal early on. But this is an investment. Ten minutes building a template means 30-second reuses forever after.

❓ I don't know what to put in [CONTEXT]

Filling two of four slots is enough. No business yet? "I'm an office worker exploring side-hustle ideas" works. A rough real context beats a polished fake one.

❓ I added a role but answers are still bland

Role too general. "Marketing expert" → "10-year B2C food brand performance marketer." More specific, more detailed.

❓ Output is long but lacks substance

You skipped the length constraint. Add "in 500 words." In 2026 models follow length instructions reliably — they just need you to state one.

❓ Same prompt, different answers each time

Normal — LLMs generate probabilistically. Weaponize it: ask for "three different directions," or rerun 2–3 times and cherry-pick. This is "generate, then curate."

❓ Which AI should I use — Claude Sonnet 4.6 vs GPT-5.4 vs Gemini 3?

All three handle the three-block prompt equally well. Differences are minor. Claude Sonnet 4.6 excels at long instruction-following and role consistency; GPT-5.4 handles image + text multimodal tasks; Gemini 3 integrates cleanly with Google Workspace. Start with whichever account you already have.

NEXT — What comes next

Twenty-five minutes ago you might have been sending "tell me how to market" and getting textbook bullets back. Now you can structure any request into context + instruction + format and pull a dramatically different result from the same AI. Over three, six, twelve months, this becomes the permanent difference between "person who works well with AI" and "person who tried AI once and didn't like it."

This is the second chapter of the SPIN series (Shape phase). The path continues:

1. Pick a tool — Pick Your First AI Tool ✓
2. Write prompts ✓ (this tutorial)
3. Ping-pong — refining thinking through dialogue (tutorial in prep)
4. Save context — Make Your Own CLAUDE.md

The best way to use this: for the next week, structure every prompt you send into the three blocks. The first three days feel slow. By day four it starts sticking. By day seven you don't even write [CONTEXT] anymore — the order just happens. Good prompting is a habit, not a talent.