A Claude Code skill that crafts well-structured, copy-paste-ready prompts for any LLM — Claude, GPT, Gemini, Perplexity, local models, image/video generators, or any other.
When invoked (e.g., "build me a prompt for..."), it guides Claude through a 5-step process:
- Understand — goal, target LLM, audience, reusability
- Gather — raw material, context, constraints
- Build — compose using 6 building blocks, format for target LLM
- Sharpen — cut fluff, be specific, ensure self-contained
- Deliver — fenced code block ready to copy-paste
npx skills add https://github.com/sungjunlee/prompt-builder.git --global --yes/plugin marketplace add https://github.com/sungjunlee/prompt-builder.git
/plugin install prompt-builder@prompt-builder
cp -r skills/prompt-builder ~/.claude/skills/prompt-builderskills/prompt-builder/
├── SKILL.md ← Core (always loaded)
├── guides/ ← Platform tips (on-demand)
│ ├── claude-guide.md ← Claude Code / web / API
│ ├── gpt.md ← GPT-5.4 / ChatGPT
│ ├── gemini.md ← Gemini 3 / Imagen 4
│ ├── perplexity.md ← Perplexity
│ └── local-models.md ← Llama 4 / Qwen / Mistral / DeepSeek
├── references/ ← Deep-dive guides (on-demand)
│ ├── components-guide.md ← 6 building blocks with examples
│ ├── prompt-patterns.md ← 8 common patterns
│ └── quality-checklist.md ← Quality checks + failure modes
└── templates/ ← Special-case templates (on-demand)
├── session-handoff.md ← AI session handoff
├── image-gen.md ← Image generation
├── video-gen.md ← Video generation
└── system-prompt.md ← Chatbot/agent system prompts
| Target | Guide | Key approach |
|---|---|---|
| Claude (Code / web) | guides/claude-guide.md |
XML tags, outcome-focused |
| GPT (5.4 / ChatGPT) | guides/gpt.md |
XML or Markdown, Thinking mode |
| Gemini (3 / 3.1) | guides/gemini.md |
XML or Markdown, few-shot preferred |
| Perplexity | guides/perplexity.md |
Search-oriented framing |
| Local models | guides/local-models.md |
Plain text, short prompts |
| Image generation | templates/image-gen.md |
Flat text, negative prompts |
| Video generation | templates/video-gen.md |
Cinematic verbs, style-first |
- No named frameworks — CO-STAR, RISEN, etc. are training wheels. 6 building blocks cover all of them.
- XML as universal default — all 3 major providers recommend it as of 2026.
- Context over instruction — 400 tokens of context + 100 of task beats 400 of instruction + 100 of context.
- Less is more — cut verbose roles first, then restated context, then hedging. Never cut examples or success criteria.
v0.5.0 — March 2026. Updated for Claude Opus 4.6, GPT-5.4, Gemini 3.1, Imagen 4, Midjourney V7, Sora 2, Veo 3.1, Seedance 2.0.