Sample-Driven AI Output: From Cant Describe It to Get It At a Glance

A Real Frustration

I wanted to make a PPT about the "Flywheel Effect."

The first time, I told AI directly:

"Help me make a PPT about the flywheel effect. It should be professional, clean, and impressive."

AI generated it quickly. I opened it — garish colors, chaotic layout, inconsistent font sizes, content piled together like a blob of paste.

I gave feedback and asked AI to revise. Three rounds of changes, each with new problems. Eventually I gave up — it wasn't that AI couldn't do it, it was that I couldn't clearly describe what I wanted.

The Turning Point: Give a Sample

Later I tried a different approach.

I found a PPT screenshot online that I thought looked great, sent it to AI, and said:

"I want a PPT in this style. Color scheme reference this, layout reference this, font pairing reference this. The topic is the Flywheel Effect, covering the physics definition, management applications, and enterprise cases."

Then I added some detail requirements:

"Title slide: dark background + white text. Content slides: card-style layout with shadow effects. Overall style: professional and clean, no flashy animations."

This time, the PPT AI generated satisfied me at a glance.

Same AI, same requirements — why such a huge difference?

The Brain's "Fuzziness" vs The Eyes' "Precision"

This experience made me realize a key problem:

What the brain imagines is fuzzy, but what the eyes see is specific.

When you use words to describe "a clean and impressive PPT," AI's understanding of "clean" and "impressive" is completely different from what's in your head. Everyone interprets those words differently, and AI can only guess based on its own understanding.

But when you show it a sample directly, AI can precisely capture:

Color scheme: Primary #1a3a5c, secondary #4a90d9, background #f5f7fa
Font pairing: Title 44px bold, body 18px regular
Layout: Card-style with shadows, 16px border radius
Information density: 2-3 cards per slide, 2-3 points per card
Visual hierarchy: Title → Card title → Card content → Page number

These things would take 100 words to describe, and AI might still not get it right. But with one screenshot, AI understands immediately.

The "Give a Reference" Technique

I summarized this discovery as a technique: "Give a Reference."

When making requests, give AI one or more actual samples and tell it "I want something like this" — it's 10x more effective than describing in words.

This technique applies to nearly all AI output scenarios:

Design Tasks

PPT creation: Provide reference PPT screenshots or HTML files
Poster design: Provide reference poster images
Web design: Provide reference website screenshots or URLs
Logo design: Provide reference logo images

Writing Tasks

Article writing: Provide a reference article, say "I like this writing style"
Copywriting: Provide reference copy, say "Write in this style"
Email writing: Provide a reference email, say "Tone reference this"

Code Tasks

Component development: Provide reference code, say "Structure reference this"
Page development: Provide a reference page screenshot, say "Layout similar to this"
API design: Provide reference API docs, say "Style reference this"

Data Tasks

Report creation: Provide a reference report template
Chart generation: Provide reference chart styles
Data analysis: Provide a reference analysis report

Case Study: Complete PPT Creation Workflow

Let me walk through that successful PPT creation process to show the full application of the "Give a Reference" technique.

Step 1: Prepare the Sample

I found an HTML-formatted PPT sample online whose design style I liked:

Blue and white color scheme, professional and clean
Card-style layout with shadow effects
Semantic HTML5 tag structure
Viewable directly in browser or printable to PDF

Step 2: Give Sample + State Requirements

I sent the sample file to AI and said:

"I want a PPT in this style. Topic: Flywheel Effect."

Step 3: Add Details

After AI generated the first draft, I added detail requirements:

"Title slide: dark background + white text. Content slides: card-style layout. Overall style: professional and clean, no flashy animations."

Step 4: Iterate and Optimize

Based on my feedback, AI generated a complete HTML PPT. I opened it in the browser — great results, only minor tweaks needed.

Total time: 10 minutes.

The previous "direct description" approach took over an hour and still didn't work.

Advanced Sample Usage

Multiple Sample Combination

You can give AI multiple samples and let it combine the strengths of each:

"Color scheme reference sample A, layout reference sample B, font pairing reference sample C."

Samples + Anti-Examples

Besides showing "what you want," you can also show "what you don't want":

"Style reference this sample, but don't pile up text densely like that other sample."

Partial Reference

You don't need to provide a complete sample — you can reference just a specific part:

"The card shadow effect reference this, but change the color to blue."

Why Is This Technique So Effective?

Explained from how AI works:

AI's essence is "predicting the most likely output based on context." The more precise the context you provide, the more accurate its prediction.

Text description: AI needs to first translate text into "visual concepts," then generate output. Two conversions, information loss at each step.
Direct sample: AI extracts visual features directly from the sample, skipping the "translation" step, dramatically reducing information loss.

It's like communicating with a designer:

"I want a clean, impressive design" — the designer can only guess
"I want a design in this style" (while showing a reference image) — the designer understands immediately

Important Notes

1. Sample quality determines output quality. Give AI a bad sample, and it will produce bad output. The sample is AI's "ceiling" — it's hard for it to exceed the sample's quality.

2. Sample ≠ plagiarism. Giving a sample tells AI "I want this style," not to copy the content. AI will use the sample's style to recreate based on your topic and content.

3. Detail requirements still matter. The sample solves the "style" problem, but "content" and "details" still need your control. After providing a sample, you still need to add specific content requirements and constraints.

Summary

The core of the "Give a Reference" technique:

Let your eyes speak, not your mouth — giving samples is 10x more effective than describing
The sample is the ceiling — choose high-quality samples
Detail requirements still matter — the sample handles style, you handle content
Applicable to all scenarios — design, writing, code, data all work

Next time before you ask AI to create anything, spend 2 minutes finding a reference sample. This habit will double your AI productivity.

Series Articles:

The reference technique described in this article is deceptively simple, but its impact on the quality of AI output is profound. The reason it works so well is fundamental to how large language models operate: they are pattern-matching engines that excel at imitating concrete examples but struggle with abstract descriptions. "Clean and professional" means nothing to an AI — it has no shared understanding of those adjectives. But a specific screenshot with specific colors, fonts, and spacing gives the AI a concrete pattern to replicate. Once you internalize this principle, you will find yourself using it everywhere: providing a sample email before asking AI to write one, sharing a code snippet before asking AI for a new function, or showing a chart before asking AI for a data visualization. Each time, the result will be dramatically better than a text-only description. The two minutes you spend finding a reference will save you the thirty minutes you would have spent correcting a mediocre result. One advanced application of this technique that I have found particularly powerful in professional settings is building a "reference library" — a curated collection of high-quality examples for each type of output you regularly request from AI. For instance, if you frequently generate marketing copy, keep a file of three to five exemplary pieces that match your brand voice across different channels. When starting a new brief, select the most relevant reference and watch how much closer the AI's first draft lands compared to working from a text description alone. This approach compounds over time because you can also add the best AI outputs back into your reference library, creating an ever-improving feedback loop that systematically raises the floor on output quality.

The reference technique works because it leverages AI pattern matching at its strongest point: reinterpreting concrete visual or structural patterns rather than decoding abstract verbal descriptions.

Ultimately, the reference technique endures because it aligns with what large language models do best. While humans can effortlessly extract stylistic principles from a single example and apply them to entirely new domains, AI systems excel when given concrete patterns to reinterpret. This asymmetry is not a limitation but a design insight: by meeting AI where its capabilities peak, you get dramatically better results with less effort than trying to articulate in words what your eyes already know.