DEV Community: Dhaval Prajapati

How to Build an AI Presentation with Claude + Gamma (Step-by-Step Guide)

Dhaval Prajapati — Tue, 04 Aug 2026 07:35:55 +0000

Spending 6+ hours tweaking PowerPoint alignment grids is a waste of creative energy — whether you're a student prepping a final-year project or a founder pitching a client.

Here's the exact workflow I use to go from a blank prompt to a polished, presentation-ready deck in minutes, using Claude AI + the Gamma connector.

Quick summary

No design skills required — Claude's reasoning + Gamma's slide engine do the heavy lifting
1-minute connector setup, no API keys needed
A battle-tested 12-slide prompt template you can copy and reuse
Bonus: feed Claude your own docs/PDFs to turn existing content into a deck automatically

Step 1: Nail down your topic first

Before touching any AI tool, know exactly what the presentation is for. A final-year engineering project needs a completely different structure than a B2B sales pitch or an internal training deck.

If you don't have a clear outline yet, brainstorm talking points with Claude or ChatGPT before generating slides.

Lesson learned the hard way: AI can build the deck, but it can't answer tough questions when someone in the room pushes back. Know your material well enough to explain it without slides — that's also what lets you catch and fix bad AI output instead of blindly accepting it.

Step 2: Connect Gamma to Claude

Open your workspace at claude.ai
Open the Connector Search menu → Add Connector → Browse Connectors → search Gamma
Click Connect, authenticate with Google SSO or email, and hit Allow

That's it — the connector is live.

Step 3: Use this prompt template

Generic prompts like "make a presentation on AI" produce generic walls of text. This is the exact structure I use — copy it and swap in your topic, audience, tone, and slide count:

Use Gamma to generate a PowerPoint presentation on [TOPIC].

Purpose: [educate / pitch / train / report]
Audience: [beginners, business owners, students, clients]
Slide count: [10-14]
Tone: [professional / conversational / persuasive]
Theme: [Minimal / Corporate / Dark Mode / Editorial / Vibrant]

Structure:
1. Title slide — topic + subtitle hook
2. Introduction — 1 attention-grabbing stat/question
3. Agenda/overview — bullet map of what's covered
4-7. Key points — core idea + supporting facts/data each
8. Comparison/before-after or chart slide
9. Common mistakes / challenges — 3-4 bullets
10. Best practices / action steps
11. Conclusion — recap + memorable closing line
12. Call to action

Content rules per slide:
- Max 5 bullets, 6-8 words each
- One core idea per slide
- Use numbers/stats over vague claims
- Speaker notes: 2-3 sentence expansion per slide

Design rules:
- Icons instead of bullet dots
- One chart slide per data-heavy section
- 2 primary + 1 accent color, consistent throughout
- High-contrast text, one font pair
- Full-bleed images only for section breaks

Avoid: walls of text, generic stock photos, more than 2 fonts, cluttered slides.

Why this works:

Strict bullet caps (5 bullets, 6-8 words) force high-density insight over text walls
Speaker notes separation keeps your spoken narrative distinct from on-screen text
Locked-in design rules stop Gamma from generating inconsistent, cluttered layouts

Step 4: Check the deck manually

AI models make mistakes — every time. Budget 5-10 minutes for a manual pass:

✅ Fact-check every number, date, and claim
✅ Confirm the slide flow reads naturally
✅ Check contrast on both laptop and projector
✅ Test-export to .pptx or PDF if you'll need offline reliability

Bonus tip: feed it your own docs

Already have a project doc, design spec, or quarterly report written up? Upload it to Claude before running the prompt, and tell it:

"Use the attached project documentation as the exclusive factual reference for generating this Gamma presentation deck."

I did exactly this for my final-year college presentation — uploading my technical docs saved me from re-explaining context Claude would've otherwise had to guess at, and produced accurate, tailored slides instantly.

FAQ

Can I export to PowerPoint?
Yes — once Gamma processes the prompt via Claude, export to .pptx, PDF, or present interactively online.

Do I need a paid subscription?
The connector itself works on a standard account. Heavy deck generation uses Gamma AI credits, and large document uploads work best with Claude Pro.

Can I upload PDFs or Word docs?
Yes — attach PDFs, markdown, or text files directly in Claude. It extracts key claims and data to populate slides accurately.

Why does topic knowledge still matter if AI builds the slides?
AI handles formatting and initial structure. Live Q&A still requires deep domain understanding — that's what lets you catch hallucinations and defend your ideas.

Originally published on dhaval-prajapati.onrender.com. If you found this useful, check out my guide on writing effective AI prompts or my case study on building a privacy-first ATS resume analyzer.

Building a Privacy-First ATS Resume Analyzer with Next.js & BYOK AI

Dhaval Prajapati — Mon, 20 Jul 2026 09:56:22 +0000

Built Resume Intelligence with one rule: no backend, no database, resume never leaves the browser.

How it works:

PDF parsed 100% client-side via Web Workers (pdfjs-dist)
BYOK — bring your own Gemini/Groq/Mistral/OpenAI key, browser calls the provider directly
Strict JSON schema forces consistent AI output across providers: score, flagged lines, missing keywords, bullet rewrites
Bounding-box PDF inspector — click a flagged issue, it scrolls + highlights the exact line
Past scans stored in IndexedDB, not a server

Full build breakdown + copy-paste master prompt: https://dhaval-prajapati.onrender.com/blog/how-i-built-ats-resume-analyzer-blog

Try it free: https://ats-resume-analysis-sigma.vercel.app/

I Gave Claude, GPT-4o, and Gemini the Same Landing Page Prompt — Here's Who Won

Dhaval Prajapati — Fri, 17 Jul 2026 17:36:44 +0000

I ran the exact same web dev prompt through three AI models — Claude, GPT-4o, and Gemini — and asked each to build a complete landing page from a single index.html file. Same prompt, zero edits, three very different results.

The Test

The prompt asked for a full landing page for a fictional tuition company, requiring:

Sticky nav with mobile hamburger menu
Hero section with SVG illustration + stats
Courses grid, testimonials carousel (auto-slide + dots)
Real-time JS form validation
Scroll animations via Intersection Observer
Zero frameworks — no Bootstrap, no Tailwind, no jQuery

Results

Feature	ChatGPT	Claude	Gemini
Lines of code	99	1,157	1,915
Overall score	52/100	🥇 89/100	🥈 84/100
Typography	System default	Playfair + DM Sans	Playfair + Plus Jakarta Sans
Scroll animations	❌ None	✅ Full staggered	✅ Full scroll reveal
Form validation	Alert only	✅ Real-time	✅ Inline error fields
SEO meta tags	Basic	Good	✅ Full OG setup
Prompt coverage	~30%	~85%	~95%

🥇 Claude — Best Overall (89/100)

Strongest aesthetic output: elegant hero with a gradient dot pattern, clean Playfair Display + DM Sans pairing, staggered scroll animations, and a testimonials carousel with real-time validation (including an Indian phone number regex). Minor miss: the hero stats row was thinner than the prompt asked for.

🥈 Gemini — Most Complete (84/100)

Longest and most literal output — nearly full prompt coverage, complete Open Graph/SEO metadata, and touch-swipe support on the carousel. Trade-off: verbose code that's harder to edit, and less visual cohesion than Claude's output.

🥉 ChatGPT — Basic Prototype (52/100)

Just 99 lines — missed animations, the carousel, form validation, and custom fonts entirely. Looks like it hit an output length limit rather than a capability gap.

Takeaway

Better prompting only pays off if the model can actually execute on long, detailed instructions. Claude wins on design polish and code cleanliness, Gemini wins on raw completeness and SEO scaffolding, ChatGPT is fine for a quick prototype but not a full build.

Full write-up with live interactive previews of each AI's actual output: read it here

How Search Engines Actually Work: URL Discovery Crawling Indexing Ranking (2026 Guide)

Dhaval Prajapati — Fri, 17 Jul 2026 17:34:40 +0000

Google doesn't scan the live internet every time you search — it queries a pre-built index. Here's the 7-step pipeline that gets a page from "just published" to "ranking on page 1."

The chain: URL Discovery → Crawling → Rendering → Indexing → Query Understanding → Ranking → Serving Results

1. URL Discovery

Google can't crawl what it doesn't know exists. Pages get discovered via:

Internal links (from other pages on your site)
External links (from other sites)
XML sitemaps (submitted via Google Search Console)

Quick distinction devs often mix up: a sitemap tells crawlers where to go, robots.txt tells them where not to go.

2. Crawling

Googlebot fetches the URL. Crawl frequency depends on site authority, update cadence, and your crawl budget — basically how many pages Google's willing to crawl on your site in a given window.

3. Rendering & Processing

Google renders HTML/CSS/JS using an evergreen Chromium engine. Block critical JS and your interactive content may never get parsed — a common SEO bug in JS-heavy frontend apps.

4. Indexing

Crawling ≠ indexing. Crawling reads the page; indexing decides whether to store it. If a page isn't showing up in search, check your Search Console indexing report — it could be a noindex tag, thin content, or duplicate content issue.

5. Query Understanding

NLP interprets what the user actually wants — informational, navigational, or transactional intent — before ranking even starts.

6. Ranking

No definitive "200 factors" list actually exists in practice, but relevance, page speed, and backlink quality (not quantity) are consistently the heavy hitters.

7. Serving Results

The SERP assembles featured snippets, People Also Ask boxes, and knowledge panels — increasingly optimized for zero-click answers.

Myths, Busted

❌ Sitemap submission guarantees ranking → it only aids discovery
❌ More backlinks = better rank → relevance/context matters more than volume
❌ Google searches the live web → it searches its stored index

FAQ

Can a page rank without being indexed?
No — indexing is a hard prerequisite for ranking.

How long does indexing take?
Hours to weeks, depending on site authority and technical health.

Originally published on my blog, where I write about AI, SEO, and web dev.

How to Use Google Stitch with ChatGPT + Claude to Build a Design System in Minutes

Dhaval Prajapati — Thu, 11 Jun 2026 04:51:31 +0000

Most people open Google Stitch and type "a modern website for my startup."

The output is fine. It's also completely generic.

The fix? A 3-AI pipeline — and it takes under 5 minutes.

The Workflow at a Glance

ChatGPT → brainstorm your design brief (pages, tone, audience, layout)

Claude AI → refine it with hex codes, typography specs, WCAG contrast, component descriptions

Google Stitch → paste the refined prompt, get a multi-screen UI prototype + React export

Each tool has one job. Stitch is only as good as the prompt you give it — preparing through two LLMs before rendering is what separates premium output from generic noise.

Why It Works

What you give Stitch	What you get
"Dark blue, professional"	Generic navy layout
`#4f46e5 primary, #0ea5e9 accent, Inter 400/600, 12-col grid, 80px section padding`	On-brand, production-ready mockup

Claude converts vague descriptions into design-system-precise language. Stitch responds to specificity — dramatically.

Quick Start Prompt (Claude step)

Here's my ChatGPT design brief: [PASTE]

Improve it for Google Stitch:
1. Replace color descriptions with specific hex codes (WCAG AA compliant)
2. Add font family, weights, base size, line-height
3. Specify grid, padding, border-radius values
4. Describe key components precisely
Output as one detailed paragraph.

Pro Tips (Quick)

🔍 Paste a competitor URL into Stitch → extract their design system in 45 seconds
📱 Always specify device target (mobile-first or 1440px desktop)
💾 Save your refined Claude prompt in DESIGN.md for reuse across pages

Want the Full Guide?

This post covers the essentials — but the full article includes:

✅ Complete ChatGPT brainstorming prompt template

✅ Full Claude refinement prompt template

✅ Live multi-screen Scoobies e-commerce prototype (interactive)

✅ Complete Stitch prompt (800+ words, production-ready)

✅ Zoom-Out-Zoom-In technique explained

✅ ChatGPT vs Claude vs Stitch comparison table

✅ From Stitch to live site (React → Vercel pipeline)

✅ 7 FAQ answers

👉 Read the full guide on my blog

By Dhaval Prajapati — AI Web Developer & Digital Marketer, Ahmedabad.

How to Write the Best AI Prompt in 2026 (No Prompt Engineering Course Needed)

Dhaval Prajapati — Sat, 30 May 2026 08:44:34 +0000

Most people say AI is overhyped.

They're wrong. They just have a prompts problem.

You know what you want — but you can't put it into words the AI actually understands. So you type something short. ChatGPT gives back something generic. You tweak it three times. Still not right. Eventually you just do the task yourself and think "AI is useless."

That's not an AI problem. That's a prompts problem. And the fix isn't complicated — you don't need a prompt engineering course. There's a simpler system.

Why Most People Get Poor AI Results

Tools like ChatGPT AI and AI Claude are built on LLMs — Large Language Models — developed by companies like OpenAI and Anthropic. These models are trained to respond to natural language inputs. Plain human language, not code, not commands — just words.

But when your natural language inputs are vague, the LLM has nothing solid to work with. It guesses. And guesses are always generic.

A ChatGPT prompt that actually works gives the LLM four things. Most people give it one. That's why results feel off.

🎯 Element	❓ Question to Answer
What	What do you want?
Who	Who is it for?
Format	What format / length?
Context	What's the background?

Cover all four and your AI output improves immediately. The problem is most people don't know how to fill all four before they start typing. That's what this method solves.

What Is Prompt Engineering — And Why You Don't Need to Learn It

Prompt engineering is the practice of writing clear, structured natural language inputs to get the best possible output from a large language model. It's a real skill. OpenAI and Anthropic both publish guides on it. There are full prompt engineering courses on every learning platform.

But here's the honest truth — you don't need any of that for regular work.

Learning AI prompt engineering properly takes time. For someone who just wants better results from AI ChatGPT or AI Claude today, a prompt engineering course is overkill.

What you need is a system that handles the hard thinking for you. And the funny thing? The LLM itself can build that system. That's the whole trick.

The 2-Tool Method — Use AI to Write Your Prompt

Simple idea — use AI to write your prompt before you use AI for the actual task.

Two tools. ChatGPT AI first, then AI Claude.

✏️ Step 1 — Figure Out What You Actually Want

Before opening any tool, write your task in plain language. Not "I want a blog post" — more like "1000-word blog post for people new to AI, friendly tone, one practical tip."

No tools yet. Just getting clear on what you're asking for.

✏️ Step 2 — Write Down the Key Points

Two minutes. Bullet points.

Audience — who is this for?
Length — how long?
Format — bullet points, paragraphs, table?
Tone — professional, casual, friendly?
Things to include — must-haves
Things to skip — what to avoid

🟢 Step 3 — Ask ChatGPT AI to Build the Prompt

Open ChatGPT AI (by OpenAI) and say:

"I need to [your task]. Here are my notes: [key points]. Write me a detailed prompt I can use with an AI tool."

ChatGPT AI turns a messy pile of ideas into something complete. Covers angles you missed. Builds structure you didn't think of.

After this step — a real, complete ChatGPT prompt.

🟣 Step 4 — Use Claude as Your Prompt Polisher

Copy what ChatGPT AI gave you. Open AI Claude (by Anthropic) and say:

"Here's a prompt I'm going to use. Make it clearer and more precise. Remove anything vague."

Claude doesn't just reorganize — it tightens. Language gets sharper. Vague parts get replaced with specific instructions. The LLM now has zero room to misinterpret.

✅ Step 5 — Run It

That's it. First-try output that actually matches what you wanted.

Why This 2-Tool Approach Works

ChatGPT AI and AI Claude are both powerful LLMs — but they think differently. Use their strengths together.

	ChatGPT AI (OpenAI)	AI Claude (Anthropic)
Strength	Brainstorms wide	Thinks precisely
Good at	Covering every angle	Removing vagueness
Result	Nothing is missing	Nothing is fuzzy
Best for	Structure	Refinement

Use ChatGPT AI alone → complete but slightly vague
Use AI Claude alone → sharp but sometimes missing something
Use both in order → nothing missing, nothing vague ✅

Real Proof — Live Example Walkthrough

I used this method to build a complete landing page for an offline tuition company.

What I Asked ChatGPT AI (Step 3)

Write a prompt to create a simple landing page for a company that provides offline tuition classes. Build using HTML, CSS, and JavaScript. Design should be clean, modern, responsive. Include hero banner, about section, courses, benefits, testimonials, contact form, and a strong call-to-action.

ChatGPT gave me a solid prompt — but read the output carefully:

"Design the page with a clean educational theme, **modern typography, **appealing colors, **soft shadows, and easy navigation."

Words like "appealing colors" and "modern typography" are still vague. An LLM filling in those gaps will just guess. This is exactly where AI Claude comes in.

The Before & After — Side by Side

❌ ChatGPT AI output (vague):

"Design the page with a clean educational theme, modern typography, appealing colors, soft shadows, and easy navigation."

✅ Claude output (precise):

"Use a professional color palette built around **deep blue, white, and soft accent colors like green, gold, and violet. Choose distinctive and elegant typography — **avoid generic system fonts* — and pair a stylish display font with a refined body font imported from Google Fonts. Use rounded cards, soft box shadows, smooth transitions, and subtle background patterns to create a premium and polished visual identity."*

Every vague word got replaced with a specific, actionable instruction. The LLM now has zero gaps to fill with generic guesses.

The Final Output

✅ First try. Three files — index.html, style.css, script.js — exactly as specified. Zero revisions.

📦 Section	What Got Built
Hero Section	Dark navy theme, bold headline, Enroll Now + Book Free Demo CTAs, SVG illustration, stats badge
About Section	Company intro, three teaching pillars, years of experience badge
Courses Grid	6 course cards — Maths, Science, English, Computer Basics, Competitive Prep, Social Studies
Contact Form	Two-column layout, real-time JS validation, success popup on submission
Testimonials	Auto-sliding carousel, star ratings, prev/next controls and dot indicators
Fully Responsive	Mobile-first design, hamburger menu, scroll-triggered animations, back-to-top button

Comparison: All Prompting Methods

Method	Time	Output Quality	Skill Needed
Random / vague prompts	1 min	❌ Low	None
Copying prompts from Reddit/Twitter	5 min	⚠️ Hit or miss	None
Taking a prompt engineering course	Weeks	✅ Very High	High
Learning full AI prompt engineering	Months	✅ Very High	Very High
⭐ This 2-tool method (ChatGPT + Claude)	10–15 min	✅ High	None

Tips to Get Even Better Results

🎯 Get specific in Step 1 — the clearer you are before touching any tool, the better everything else goes. Vague task in = vague prompt out.

👥 Always mention who the output is for — audience is the most powerful context you can give a large language model.

📐 Name the format explicitly — bullet points, paragraphs, table? Tell the AI. Don't leave it to guess.

🔍 If the result still feels off — ask Claude: "What's missing or vague in this prompt?" It'll tell you exactly what to fix.

💾 Save your good prompts — Notion, Notes, Google Doc. Build a personal library. You will use them again.

Try It on Your Next Task

Most people read something like this, think it's useful, and do nothing. Next time you open any AI tool — run this method.

Step	Action
1️⃣	Write your task in plain language
2️⃣	List key points (audience, tone, format)
3️⃣	Build the prompt in ChatGPT AI
4️⃣	Polish it in AI Claude
5️⃣	Run it

No prompt engineering course. No special skills. Just a smarter system.

💡 If this helped — share it with someone who keeps saying AI never gives them what they want.

About the Author

Dhaval Prajapati — AI Web Developer & MCA Graduate

I build AI-powered web experiences and write about practical AI workflows for developers.

🌐 Portfolio: dhaval-prajapati.onrender.com
📝 Blog: dhaval-prajapati.onrender.com/blog.html
💼 LinkedIn: linkedin.com/in/dhaval-prajapati-900b8818b
🐙 GitHub: github.com/dhavalMCA

Originally published on my portfolio blog → dhaval-prajapati.onrender.com/blog-single.html?id=b1