DEV Community: ZHANG YICHI

How I built a free, open-source library of AI slide-deck prompts (120+, by use case)

ZHANG YICHI — Thu, 15 Oct 2026 07:29:04 +0000

Every time I needed a deck, I'd open an AI image model and rewrite the same "make me a pitch deck" prompt from scratch — and the output was hit-or-miss. So I started saving the prompts that reliably produced good slides, and eventually organized 120+ of them by use case (pitch, QBR, research, sales, education) and visual style (minimal, dark, corporate, editorial). Every prompt has a preview image.

It's free and open-source. Here are three that consistently work:

**1. Minimal SaaS pitch deck

Create a 10-slide minimal SaaS startup pitch deck. White background, electric blue #0066FF accent. Slide 1: bold product name + one-line tagline, logo placeholder top-left. Slide 2: Problem — 3 pain points with thin line icons. Slide 3: Solution — UI mockup placeholder, 3 labeled callouts. Slide 4: Market Opportunity — TAM/SAM/SOM nested circles with $ values. Slide 5: Business Model — 3-column pricing table, recommended tier highlighted with a badge. Slide 6: Traction — MoM active users line chart, milestone markers. Slide 7: Competitive Landscape — 2x2 positioning matrix vs 4 named competitors. Slide 8: Team — 4 headshot cards, name/title/one-line bio. Slide 9: 3-Year Revenue Forecast — grouped bar chart, CAGR annotation. Slide 10: Close — full-bleed CTA, funding ask amount, contact email. Font: Inter throughout. Consistent header bar with logo + slide number on every slide. Tone: confident, data-driven, investor-ready.

Why it works: it names every slide's job (problem, solution, TAM/SAM/SOM, traction, ask) and locks one accent color + one font, so the 10 slides read as a system instead of 10 unrelated images.

**2. Dark-mode dev-tool launch keynote

Design a 12-slide dark-mode developer tool launch keynote. Background #0D0D0D, neon green #00FF87 accent. Slide 1: full-bleed hero, product name in JetBrains Mono. Slide 2: Problem — terminal-style error messages as bullet points with red # markers. Slide 3: Product reveal — large screenshot with glowing border and subtle particle effect placeholder. Slides 4-7: four feature slides, alternating code-snippet left / result screenshot right layout. Slide 8: Architecture diagram — dark nodes, luminous connecting arrows, system labels in monospace. Slide 9: Benchmark bar chart — dark bg, green bars vs grey competitor bars. Slide 10: Pricing — 3 dark tier cards, Pro highlighted. Slide 11: GitHub star + Discord join CTA. Slide 12: Launch date countdown closer, full-screen gradient. Font: JetBrains Mono for code elements, Inter for prose. Tone: authoritative, technical, exciting.

Why it works: the alternating code-left / screenshot-right rhythm and a single neon accent give it a real keynote cadence; monospace for code vs. Inter for prose keeps hierarchy clean.

**3. Marketing QBR

Build a 14-slide Marketing Quarterly Business Review (QBR) deck. Corporate style: white background, bold teal #00BFA5 accents, dark grey header bars. Slide 1: Executive Summary — 6-KPI dashboard cards (MQL, SQL, CAC, LTV, Revenue, NPS). Slide 2: Channel Performance — heat map table, red/amber/green coding. Slide 3: Campaign ROI table sortable by spend. Slide 4: SEO & Organic Traffic — waterfall chart showing channel contributions. Slide 5: Social Media Engagement — dual-axis trend chart (reach vs. engagement rate). Slide 6: Paid Ads — ROAS by channel stacked horizontal bar. Slide 7: Email Funnel — conversion funnel diagram. Slide 8: Competitive Share-of-Voice — pie chart comparison. Slide 9: Wins & Learnings — 2-column table. Slides 10-11: Next Quarter OKRs with quarterly milestone timeline. Slide 12: Budget Request — waterfall bar. Slide 13: Team Highlights. Slide 14: Appendix placeholder. Font: Helvetica Neue. Tone: data-driven, executive-ready.

Why it works: it front-loads a 6-KPI executive summary and assigns a specific chart type per slide (heat-map, waterfall, funnel, stacked bar), which is what makes a review deck feel data-driven instead of decorative.

How to use: copy any prompt → paste into your image model or SurgePix → generate → then edit (swap a chart, retitle a slide, recolor all 12 to match) rather than regenerating and hoping.

Browse the full library (120+, with previews)

Source (open, MIT): github.com/SurgePix/ai-ppt-prompts

Contributions welcome — open a PR with a prompt + its output.

How to Turn Any Article Into an Illustration Series with Claude Code

ZHANG YICHI — Thu, 23 Jul 2026 09:50:37 +0000

I never opened a design app for this.

One prompt in Claude Code, and I had a full set of blog illustrations — a cover image plus three section illustrations — same character, same style, consistent 16:9 layout, ready to drop into a post.

Here's the exact setup, in case you want the same thing.

The problem with most AI image tools

If you've generated images with an LLM before, you know the annoying part: consistency. Ask for four illustrations on related topics, and you'll usually get four different styles, four different "main characters," sometimes four different color palettes. Fine for a single hero image. Useless for a matching set.

For a blog post that needs multiple illustrations — a cover plus a few section images — that inconsistency is a dealbreaker. Nobody wants a post where every illustration looks like it came from a different artist.

The fix: lock the style into a skill, not a prompt

Instead of re-describing the style every time and hoping the model cooperates, I used SurgePix's generate-illustrations skill for Claude Code. It's one of nine open-source agent skills that let coding agents (Claude Code, Cursor, Codex CLI, Gemini CLI) call SurgePix's image API directly — no browser, no dashboard, just natural language in your terminal.

The skill handles the consistency problem at the generation level: same character, same linework, same palette, across every image in a set — because the style constraints are baked into how the skill calls the API, not left to chance in a one-off prompt.

Setup (3 minutes)

1. Get your API key

Go to surgepix.ai → Settings → API Management → + New Key.

2. Install the skill

npx skills add SurgePix/agent-skills

3. Set your API key as an environment variable

Windows (PowerShell):

[Environment]::SetEnvironmentVariable("SURGEPIX_API_KEY", "your-key", "User")

Mac/Linux (zsh):

echo 'export SURGEPIX_API_KEY=your-key' >> ~/.zshrc
source ~/.zshrc

4. Verify

node surgepix-setup/scripts/check_env.mjs

If that prints your key back, you're done. Restart your terminal/agent if it doesn't pick it up right away.

Works with Claude Code, Cursor, Codex CLI, Gemini CLI, and any agent that supports the SKILL.md standard.

Generating the set

Once it's installed, you don't need to remember any commands — just describe what you want in plain language. Here's the actual prompt I used:

Generate a set of blog illustrations for my post about how AI agents use tools via MCP: one cover, and 3 section images — "what is MCP", "how tool calling works", "security". Use the same character in all four. Consistent hand-drawn style, 16:9.

Claude Code recognized the request, loaded surgepix-generate-illustrations, checked the environment, and generated all four images in one pass — returning a download link when it finished. No manual regeneration, no style drift between images.

Why this matters beyond illustrations

This is really a pattern, not a one-off trick: agent skills let you offload an entire workflow — not just a single API call — to natural language. The same approach works for other SurgePix skills in the repo: posters, presentations (with real .pptx export, not just images), background removal, image translation, and more. If you're writing technical content and hand-drawing (or Canva-ing) your own illustrations, this is worth trying.

Try it yourself

Full illustration skill demo (video): https://youtu.be/qBTxv2CVpkM
Skills repo (open source): github.com/SurgePix/agent-skills
Full install + first-generation walkthrough (video): https://youtu.be/Hk34NWd80XE
Docs: surgepix.ai/blog/developer-guides/install-skills

If you try it and get stuck, drop a comment — happy to help debug.