You ask AI to write something creative. It hands you back the most predictable, committee-approved version of the idea imaginable. You ask it to follow specific instructions. It follows three of them and invents the rest. You try again, slightly differently, and wonder if you’re the problem.
You’re not. The prompts are.
There are two techniques to fix this: one that unlocks the creative range AI is actually hiding from you, and one that makes your instructions impossible to misread. They’re called magic prompting and markdown prompting, and they work well separately.
Together, they’re the difference between an AI agent that kind of helps and one that does exactly what you meant.
Think of this as your practical guide on how to prompt AI properly. No jargon, no “AI expert” required, just two structured prompting techniques you can use in the next ten minutes with immediately better results. These are already being used by AI consultants and internal AI teams to improve output reliability at scale.
If you want to understand the deeper infrastructure that makes AI context-aware in the first place, our breakdown of how structured AI systems use MCP for context management is a good companion read.
Before diving into the techniques, here’s what applying them looks like in practice.
Client scenario: A financial services client came to us struggling with inconsistent AI output across their marketing team, different writers were prompting the same tools differently and getting wildly different quality results. We built them a library of markdown prompt templates with variables for their most common content types: client emails, social posts, compliance disclosures, and campaign briefs. Within two weeks, their team’s average prompt quality was consistent enough that a junior writer was producing output on par with their senior copywriter. Magic prompting got layered in for the campaign work specifically, where they needed genuine creative variety rather than the first decent idea. The combination cut their revision cycles in half.
That’s the practical upside. Now here’s how both techniques actually work.
Magic prompting is a technique where you ask the AI to generate several possible responses with confidence scores before choosing one. Instead of the model jumping straight to its safest answer, it first maps out the landscape of what’s possible, then picks the best option.
It sounds simple because it is! But the results are not what you’d expect.
In October 2025, researchers from Northeastern University and Stanford published Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity, a study tested across creative writing, dialogue simulation, open-ended QA, and synthetic data generation. When you ask an AI to explicitly surface multiple options and assign probability scores to each, responses become measurably more diverse and creative, 1.6 to 2.1 times more so compared to standard prompting. For a technique that requires zero technical setup, that’s a meaningful return.
The alignment training that makes AI models safe and agreeable also makes them boring. Models are trained to converge on responses that feel correct, appropriate, and unlikely to cause problems. Researchers call this typicality bias, or the tendency to produce the most statistically average answer rather than the most interesting or useful one.
Magic prompting breaks that pattern by forcing the model to consider its options out loud before committing. It’s the difference between asking someone, “What should we do tonight?” and getting “I don’t know, whatever you want”, versus asking them to name three options before deciding. That little bit of structured deliberation produces better thinking.
Copy this and keep it somewhere useful:
Before answering, please:
Paste that before any creative request and watch what changes.
Use it when there’s no single correct answer. Creative tasks are the obvious ones, jokes, story ideas, campaign concepts, taglines, headlines. But it also works well for:
Don’t use it when you already know what you want, or when speed matters. Skip it for:
The Northeastern University and Stanford research showed a 1.6 to 2.1x increase in response diversity.
In practice: genuinely different ideas rather than variations on the same one, more surprising angles, and a higher ceiling on the best option in the set. The tradeoff is real, responses take longer and use more tokens. For high-stakes creative work, that’s usually worth it.
One caveat: results vary by model. The technique works best with capable frontier models such as ChatGPT’s latest models, Claude’s higher-tier versions (Sonnet, Opus, and the beautiful Fable), and Gemini Pro. On smaller or older models, you may see less dramatic improvement, or occasional confusion about what you’re asking for. Test it on your tool of choice before committing to it as a default.
Markdown is a simple formatting system created in 2004. It uses plain-text symbols to add structure, pound signs for headings, hyphens for bullet points, asterisks for bold or italic text, and backticks for code. You’ve seen it everywhere: GitHub, Reddit, Notion, Discord, most documentation on the internet.
You don’t need to be a developer to use it. If you can type a # and a dash, you’re already doing it. Wonder why AI uses a lot of em dashes? It’s used to reading markdown (MD) formatting.
When you give an AI a wall of text, it has to infer what’s important, what’s context, and what’s instruction. That inference process is where things go wrong.
Markdown removes the ambiguity. Clear headings tell the model what each section is for. Bullet points make instructions sequential and scannable. Bold text signals priority. Code blocks show exactly what format you expect.
Multiple studies back this up. A 2025 review in ScienceDirect found that modifying both the structure and content of a prompt has a notable influence on model behavior, and that well-constructed prompts can directly reduce hallucinations. A separate PMC-published study testing GPT-4, LLaMA 2, and DeepSeek found that vague, unstructured prompts produced hallucination rates more than twice as high as structured equivalents. And The Prompt Report, a systematic survey of 1,565 papers on prompting techniques, concluded that structured prompt design is the single most reliable lever for improving output accuracy across model types. Structured input produces structured output. That treats a probabilistic reasoning system like the tool it actually is, not a search engine you’re just throwing keywords at.
# Headings — use for your primary goal (H1), context and instructions (H2), specific details (H3). Creates clear hierarchy so AI distinguishes between main goals and sub-instructions.
– Bullet lists / 1. Numbered lists — bullets for unordered requirements, numbers for sequential steps. Makes instructions scannable and ensures the AI follows each item systematically.
**Bold** and *italic* — bold for critical requirements, italic for emphasis or terminology. Draws AI attention to priorities.
`Code blocks` — use to show precise output format or exact strings the AI should preserve. Prevents formatting errors.
Save this template. Use it every time a request has more than two moving parts:
# GOAL
[What you want to achieve in one sentence]
## CONTEXT
[Background information the AI needs]
## INSTRUCTIONS
– Step 1
– Step 2
– Step 3
## OUTPUT FORMAT
[Exactly how the response should look — length, structure, tone]
## EXAMPLES
[One or two examples of what good output looks like]
One underused technique that pairs naturally with markdown: variables. Instead of rewriting a prompt from scratch each time, use placeholder text in brackets that you swap out as needed.
# GOAL
Write a [CONTENT TYPE] for [AUDIENCE] about [TOPIC].
## TONE
[TONE — e.g. conversational, authoritative, playful]
## LENGTH
[TARGET LENGTH]
This turns a one-off prompt into a reusable template. Swap the bracketed values, run the prompt, get consistent results. It’s prompt engineering 101, and it’s the fastest way to build a library of reliable, structured prompting templates your whole team can use. If you’re looking for a prompt engineering definition that actually means something practical: it’s this, designing inputs that reliably produce the outputs you need.
Over-engineering a simple prompt is its own kind of problem.
Example 1: Email Draft Request
Plain prompt:
Write me an email to a client who missed a deadline asking them to update me without being rude. Keep it professional but personable.
AI Response:
Markdown prompt:
# GOAL
Draft a professional follow-up email to a client who missed a project deadline.
## CONTEXT
– The client is a long-term partner (3 years)
– The missed deadline affects our internal timeline by two weeks
– We want to maintain the relationship while getting a clear update
## INSTRUCTIONS
– Keep the tone warm but direct
– Ask for a specific update by end of week
– Avoid language that assigns blame
## OUTPUT FORMAT
Subject line + email body, under 150 words
AI Response:
Example 2: Content Brief Creation
Plain prompt:
Make a content brief for a blog post about email marketing for small businesses.
AI Response: (I truncated some of the responses because they were way too big. This one had 5 sections 😬)
Markdown prompt:
# GOAL
Create a full content brief for a TOFU blog post on email marketing for small business owners.
## CONTEXT
– Audience: Small business owners with 1-10 employees
– They’re likely using Mailchimp or just starting out
– Competitors are HubSpot, Klaviyo blog. We need to be more practical and less corporate
## INSTRUCTIONS
– Include target keywords, required sections, and internal linking opportunities
– Flag which sections should include examples or data
## OUTPUT FORMAT
Match the structure of a standard content brief with headings for each component
AI Response: (This was INSANELY in-depth. Again, truncated here for space purposes but it’s massive and included a ton of great research suggestions.)
Example 3: Code Generation
Plain prompt:
Write me a Python function that sends a Slack message.
AI Response:
The output actually wasn’t all that bad; it just didn’t have all of the details I needed.
Here’s the code it gave me:
import logging
from slack_sdk import WebClient
from slack_sdk.errors import SlackApiError
def send_slack_message(token: str, channel: str, text: str) -> bool:
"""
Sends a message to a Slack channel using the WebClient API.
:param token: Your Slack Bot OAuth Token (xoxb-...)
:param channel: Channel name (e.g., "#general") or Channel ID (e.g., "C1234567890")
:param text: Plain text or markdown-formatted message to send
:return: True if sent successfully, False otherwise
"""
client = WebClient(token=token)
try:
response = client.chat_postMessage(
channel=channel,
text=text
)
print(f"Message sent successfully to {channel}!")
return True
except SlackApiError as e:
# e.response contains the error breakdown from Slack
print(f"Failed to send message: {e.response['error']}")
return False
# --- Example Usage ---
if __name__ == "__main__":
# Best practice: Load your token from an environment variable rather than hardcoding!
SLACK_TOKEN = "xoxb-your-bot-token-here"
CHANNEL_NAME = "#general"
MESSAGE = "Hello from Python! 🚀"
send_slack_message(token=SLACK_TOKEN, channel=CHANNEL_NAME, text=MESSAGE)
Markdown prompt:
# GOAL
Write a Python function to send a formatted Slack message via webhook.
## REQUIREMENTS
– Use the `requests` library only (no Slack SDK)
– Accept message text and optional username as parameters
– Return True on success, False on failure with a printed error
## OUTPUT FORMAT
Include type hints and a docstring
## EXAMPLE USAGE
send_slack_message(“Deployment complete”, username=”DeployBot”)
AI Response:
Here’s the code this markdown prompt gave me:
import requests
def send_slack_message(
text: str, webhook_url: str, username: str | None = None
) -> bool:
"""Sends a formatted message to a Slack channel using an incoming webhook.
Args:
text: The message body text to send.
webhook_url: The incoming Slack webhook URL.
username: Optional custom display name for the bot.
Returns:
True if the message was posted successfully, False otherwise.
"""
payload: dict[str, str] = {"text": text}
if username:
payload["username"] = username
try:
response = requests.post(webhook_url, json=payload, timeout=10)
response.raise_for_status()
return True
except requests.exceptions.RequestException as e:
print(f"Failed to send Slack message: {e}")
return False
# --- Example Usage ---
if __name__ == "__main__":
WEBHOOK_URL = "https://hooks.slack.com/services/YOUR/WEBHOOK/URL"
# Minimal usage
# send_slack_message("Deployment complete", webhook_url=WEBHOOK_URL)
# Usage with custom username
success = send_slack_message(
"Deployment complete", webhook_url=WEBHOOK_URL, username="DeployBot"
)
print("Success:" if success else "Failed:")
This one was to the point and added everything that I required because it had explicit instructions for what I wanted. That’s the difference!
More context = better understanding.
Markdown gives the AI a clear map of your request. Magic prompting gives it permission to explore the territory before committing to a path. Together, they produce outputs that are simultaneously well-structured and creatively ambitious.
On their own, each technique solves half the problem. Markdown fixes clarity. Magic prompting fixes range. Combined, you get an AI that understands exactly what you need and brings its best thinking to filling it.
# GOAL
[Your objective in one sentence]
## CONTEXT
[Background the AI needs]
## INSTRUCTIONS
Before providing your final answer:
## OUTPUT FORMAT
[Specify exactly what the final response should look like]
## CONSTRAINTS
– [Any hard limits or requirements]
The request: campaign concepts for a new oat milk brand targeting Gen Z, across TikTok, Instagram, and out-of-home.
The magic + markdown prompt:
# GOAL
Generate marketing campaign concepts for a new oat milk brand targeting Gen Z consumers.
## CONTEXT
– Brand is new, no existing recognition
– Target audience: 18-26, values authenticity, skeptical of corporate marketing
– Budget tier: mid-market
– Channels: TikTok, Instagram, out-of-home (OOH)
## INSTRUCTIONS
Before providing your final recommendation:
## OUTPUT FORMAT
– Campaign name
– One-line concept
– Scored options in table format
– Full development of winning concept with channel-specific executions
## CONSTRAINTS
– No corporate speak or polished brand voice
– Must work with user-generated content on TikTok
– OOH execution needs to work without a logo (brand awareness phase)
The output you get from that prompt versus “give me some oat milk campaign ideas” is not a marginal improvement. It’s a different category of response.
This kind of structured, context-rich prompting is also what makes AI-generated content actually usable at a professional level. If you’re curious how AI is being applied to broader creative work, our piece on AI-generated websites and what they can and can’t do covers where the technology stands today, and where human judgment still matters.
Both techniques work in free and paid tiers. On ChatGPT Plus, you can save markdown templates inside Custom GPTs so they’re pre-loaded every time, worth setting up if you use the same prompt structure repeatedly.
Claude has some of the strongest markdown parsing available. It’s trained heavily on structured text and responds well to clear hierarchy. The Projects feature lets you save system prompts with your templates built in, so you’re not re-pasting the framework every session.
Gemini handles markdown well across its tiers. Gemini Advanced handles complex multi-section prompts reliably. The free tier handles most markdown, though very nested structures occasionally get flattened.
“Before answering, show me 3-5 options with confidence scores, then pick the best one.”
# Main heading
## Subheading
– Bullet point
**bold**
*italic*
`code`
Magic prompting → creativity, brainstorming, campaign concepts, anything where multiple valid answers exist
Markdown → complex instructions, reusable templates, multi-part requests, formatting-sensitive outputs
Both together → high-stakes requests, marketing briefs, content creation, anything where quality matters and you have a minute to prompt properly
Using markdown on simple questions. “What’s a good subject line for this email?” doesn’t need five headings. Match the structure to the complexity.
Using magic prompting when you need speed. In a fast back-and-forth conversation, the extra deliberation step slows everything down. Save it for work that matters.
Inconsistent formatting. If you use ## for instructions in one section and plain text in another, you’ve recreated the ambiguity you were trying to eliminate.
Not testing before saving as a template. Run your prompt a few times before treating it as reusable. Edge cases show up fast.
Forgetting the output format. The most common reason AI gives you a response in the wrong structure is that you never specified the right one. Tell it exactly what you want back.
A prompt engineer is someone who specializes in designing inputs that get reliable, high-quality outputs from AI models. It’s become a real job title at a lot of companies, but you don’t need the job title to use the techniques. Magic prompting and markdown prompting are accessible advanced AI prompting techniques that anyone can apply, no engineering background required. The difference between a prompt engineer and a regular user is mostly practice and intentionality, both of which you’re building right now.
No. Markdown is plain text with a few symbols. Magic prompting is a sentence you paste at the top of your request. Neither requires any technical background.
Yes. Both techniques work with free tiers of ChatGPT, Claude, and Gemini. Paid tiers handle complex markdown structures more reliably, but the fundamentals work everywhere.
You can try both in the next ten minutes. Getting consistently good results takes a few days of experimentation — mostly figuring out when to use which technique and when to skip both.
Yes, and it’s worth doing. CO-STAR gives you a proven structure for the content of your prompt. Markdown gives you the formatting. Magic prompting gives you the creative range. They’re additive, not competing.
Magic prompting uses more tokens because you’re getting multiple responses before the final one, so yes, if you’re paying per token, it adds up. Markdown doesn’t add meaningful cost. It’s just formatting.
Most people use AI the same way they Google things: toss in a question, take whatever comes back, move on. That approach treats a probabilistic reasoning system like a search engine, and then people wonder why it underperforms.
Magic prompting fixes the creativity ceiling. Markdown fixes the instruction-following. Together, they turn a capable model into one that’s actually working with you instead of around you.
The techniques take ten minutes to learn and a few days to make habitual. After that, you’ll find it genuinely annoying to go back to unstructured prompting, not because it doesn’t work at all, but because you’ll know exactly what you’re leaving on the table.
Consider this your AI prompting guide for getting started. Next up: our prompt engineering guide on CO-STAR — a structured framework that pairs naturally with everything covered here.
Ready to build AI into your marketing in a way that actually moves the needle? Get in touch with the bgood media team.
Eighty-one percent of customers research a business online before they buy anything from it. If they search for yours and find nothing, where do you think they go instead? Straight to a competitor. That’s just how people shop now. I get asked some version of “do I really need a website?” more than almost any […]
Read Me
A consumer tells their AI assistant they’re “running low on protein powder”. Without opening a browser, the assistant checks purchase history, compares prices across retailers, filters out artificial sweeteners, applies a subscribe-and-save discount, and places the order. The interaction takes seconds, with no search bar and no scrolling through pages. That scenario already exists in […]
Read Me
You’ve been there. Twenty minutes into rewriting the same prompt, getting slightly different versions of mediocre, wondering what the person next to you is doing to get AI output that’s actually usable. The answer is almost always a framework. Usually, the person who’s getting usable AI results has a system … and you don’t. That […]
Read Me
