prompt-writer

prompt-writer

The most comprehensive prompt engineering skill for AI agents. Write, review, and improve prompts for any LLM.

Install

npx skills add taradepan/prompt-writer

Or for a specific agent:

npx skills add taradepan/prompt-writer -a claude-code
npx skills add taradepan/prompt-writer -a cursor
npx skills add taradepan/prompt-writer -a copilot

What it does

4 modes: Write new prompts, improve existing prompts, review prompts, write tool descriptions.

10 model families: Claude, GPT, Gemini, Llama, DeepSeek, Mistral, Cohere, Qwen, Grok, Amazon Nova — with model-specific guidance for each.

Research-backed: 30+ papers cited. Includes findings most guides miss:

• Prompting Inversion (simpler prompts work better on frontier models)
• Think Tool pattern (+54% on complex tasks)
• Chain-of-Draft (91% accuracy at 7.6% of CoT tokens)
• Contrastive examples (35.9% to 88.8% improvement)
• Tool description quality (97% have defects — 6 essentials to fix them)

Ready-to-use templates: Copy-paste prompt skeletons for coding agents, chatbots, data extractors, content writers, multi-agent orchestrators, plus per-model templates for Claude, GPT, Gemini, DeepSeek R1, Llama, Cohere, and Nova.

Files

skills/prompt-writer/
├── SKILL.md                          # Core skill (4 modes, 10 models)
└── references/
    ├── templates.md                  # Copy-paste templates & meta-prompts
    ├── model-specific-tuning.md      # Per-model guidance (10 families)
    ├── prompt-improvement.md         # Diagnose & fix bad prompts
    ├── security-patterns.md          # Spotlighting, defense hierarchy
    ├── research-evidence.md          # 30+ papers with measured results
    └── production-prompt-anatomy.md  # Claude Code, Cursor, Manus patterns

Key features

Feature	What you get
Write mode	5-step process: context → structure → write → review → test
Improve mode	6-dimension defect taxonomy, 8 rewriting patterns, ETGPO optimization
Model matrix	Quick-reference table for all 10 model families
Templates	5 use-case templates, 7 per-model skeletons, 3 meta-prompts
Multimodal	Image/video/audio prompting guidance per model
Structured output	JSON mode setup for each model
Security	Spotlighting (<2% attack rate), defense hierarchy, OWASP Top 10
Caching	Prompt structure for 90% cost reduction
Before/after	5 model-specific transformation examples

Compatible agents

Works with any agent that supports skills: Claude Code, Cursor, GitHub Copilot, Windsurf, Codex, Cline, Continue, CodeBuddy, OpenCode, and 30+ others.

Current as of

March 2026. Covers Claude 4.6, GPT-5.4, Gemini 3.0, Llama 4, DeepSeek R1/V3.

License

MIT