👉 Try the tool here: Prompt Workshop
You do not need a revolutionary, subscription-based AI product to organize your favorite prompt or your whole prompting workflows, you just need a practical, offline-first tool that gets out of your way and solves a very specific organizational problem.
We built a central repository to keep some of our prompting techniques from getting buried in chat threads and Notion pages. Today, we are open-sourcing the tool with a CLI, a visual browser, and a selection of 82 production-ready templates.
👉 Available as an app, a CLI a local web-app, or available on dishine domain: Prompt Workshop
The app uses local browser storage, meaning it doubles as your personal prompt library. You can build, save, and reuse your prompts without ever creating an account, just bookmark the page. Need to switch computers or back up your work? I’ve added a simple JSON export/import feature. Just download your library and load it onto any other device, giving you full cross-device flexibility while keeping your data entirely in your own hands.
🔗 View the repository on GitHub (it’s open source) | Read the wiki
👉🇫🇷 Read this article in french here | Lire cet article en français ici
👉🇮🇹 Read this article in italian here | Leggere questo articolo in italiani qui
Update: the AI-Prompt Library just got more powerful. Alongside new interactive features, we’ve published an in-depth Wiki that provides a deep dive into advanced prompt engineering strategies and tool mastery. it is available in 🇬🇧 English, in 🇫🇷 French and in 🇮🇹 Italian. See the video with actual real footage of the tool:
The templates cover everything from core reasoning frameworks (like Chain-of-Thought and ReAct) to production-ready system prompts for specific roles (like Data Analyst or Executive Advisor). We also included domain-specific templates for marketing, development, data engineering, and business communication.
These are not generic “write me a blog post” prompts. They are structured templates with {{placeholders}}, examples, expert tips, and common mistakes to avoid. metadata (title, category, tags, difficulty, supported models), followed by Markdown content with sections for “When to Use”, “The Technique”, “Template”, “Examples”, “Tips”, and “Common Mistakes”.
The seven categories
The library is organized into seven distinct areas. Each prompt file follows the same structure: YAML frontmatter with metadata (title, category, tags, difficulty, supported models), followed by Markdown content with sections for “When to Use”, “The Technique”, “Template”, “Examples”, “Tips”, and “Common Mistakes”.
Frameworks (9 prompts) cover the foundational reasoning techniques that work across all models and use cases: Chain-of-Thought, Few-Shot Patterns, ReAct Agent, Tree-of-Thought, Role-Based Prompting, Meta-Prompting, Constitutional AI, Prompt Chaining, and Structured Extraction. These are the building blocks that the Compose feature draws from. If you want to improve your manual prompting, we remember you this old technique, still useful: the role reversal technique.
Model-specific (6 prompts) go deep on the quirks and strengths of individual models. The Claude guide covers XML tags, extended thinking, and 200K context strategies. The GPT guide covers JSON mode and structured outputs. Gemini covers multimodal inputs and Search grounding. There is also a comparison guide to help you choose the right model for a given task.
System prompts (6 prompts) are production-ready prompts you can drop directly into an application or API call. They cover the roles we most commonly need: Coding Assistant, Content Writer, Data Analyst, Research Assistant, Executive Advisor, and Customer Support. Each one includes behavioral rules, output format specifications, and edge case handling.
Marketing (8 prompts) cover the workflows that come up most in client work: SEO content briefs, email campaign sequences, social media calendars, competitor analysis, ad copy generation, brand voice guides, conversion copywriting, and LinkedIn content. These templates are highly effective when integrated into broader AI workflow automation engines.
Development (8 prompts) cover software engineering tasks: code review with severity classification, API design and documentation, database schema design, test generation, refactoring, architecture decision records, prompt-as-code patterns, and debugging.
Data (7 prompts) cover data analysis and engineering: SQL query building from natural language, data pipeline design, dashboard specification, data quality audits, statistical analysis, visualization specification, and ETL automation.
Business (8 prompts) cover professional communication and planning: proposal generation with MoSCoW methodology, meeting summaries, OKR generation, stakeholder updates, risk assessment, pitch deck outlines, client communication templates, and competitive intelligence.
The three ways to use it
We designed the library to fit into different workflows, depending on how you prefer to work.
The command-line interface
For developers who live in the terminal, the CLI is the fastest way to find what you need. The search command ranks results by relevance using a simple scoring system: title matches score 100 points, tag matches score 50, category matches score 30, and content matches score 10. It is not fuzzy matching, but it is fast and predictable.
The most useful CLI feature is the use command. Instead of copying a template and manually finding and replacing placeholders in a text editor, the CLI walks you through it interactively. It finds every {{placeholder}} in the template, asks you for the value one by one, renders the final prompt, and copies it to your clipboard.
The show command displays the full prompt content with metadata, including which models it works best with and what difficulty level it is. The random command shows a random prompt, which is useful for discovering templates you forgot were in the library.
The visual browser
👉 Try the tool here: Prompt Workshop
If you prefer a graphical interface, the library includes a standalone visual browser. It is a single HTML file (viewer.html) that embeds all the prompt data. You do not need to run a server; you just open the file in your browser, or run prompt-lib viewer to have the CLI generate a fresh version and open it automatically. If you don’t want to host it locally, I added it on diShine domain: https://prompt.dishine.it
The viewer gives you real-time filtering as you type, clickable tag pills to filter by topic (reasoning, copywriting, sql, etc.), and a clean Markdown rendering of the prompt content. It also includes a “Workshop” panel where you can fill in placeholders and see a live preview of your final prompt before copying it. The Workshop also lets you save filled prompts to your browser’s localStorage and export them as Markdown files.
The viewer works in both light and dark mode, and is responsive enough to use on a phone or tablet if you need to reference a prompt while working elsewhere.
Thanks to local persistence, your prompt library is securely saved on your device. This allows you to build your workspace over time, giving you full control to create, modify, export, share, and delete your custom prompts.
Programmatic import
If you are building your own AI applications, you can import the library programmatically. The loadPrompts() function parses the YAML frontmatter and Markdown content of all 52 files, returning a structured array you can use in your own code. The searchPrompts() function gives you the same relevance-ranked search as the CLI.
import { loadPrompts } from '@dishine/prompt-library';
import { searchPrompts } from '@dishine/prompt-library/src/search.js';
const prompts = loadPrompts();
const results = searchPrompts(prompts, 'code review');
console.log(results[0].title); // "Code Review Prompt"
console.log(results[0].content); // full prompt content
This is the entry point for teams that want to build their own prompt management tools on top of the library structure, or integrate the templates into a custom internal application.
The Compose feature: layering prompts
The most interesting feature in the library is the Compose function. It solves a common problem in prompt engineering: how do you combine a persona, a reasoning technique, and a specific task without creating a messy, contradictory prompt?
The Compose feature lets you build a composite prompt from three distinct layers:
- The system prompt: sets the persona, behavioral rules, and constraints (for example, “Data Analyst,” which defines the role, output format expectations, and how to handle ambiguous requests).
- The reasoning framework: adds a thinking technique (for example, “Chain-of-Thought,” which instructs the model to reason step by step before answering).
- The task template: the actual work to be done (for example, “SQL Query Builder,” which provides the specific template with placeholders for the query context).
When you select these layers in the visual browser or the CLI, the tool concatenates them cleanly under three labeled sections (# SYSTEM PROMPT, # REASONING FRAMEWORK, # TASK TEMPLATE), extracts all the unique placeholders from all three layers, deduplicates them, and presents you with a unified form to fill out.
The result is a single prompt you can copy and paste directly into your model of choice. The composed prompt is also saved to your library if you click “Save to My Library” in the viewer, so you can reuse the same combination later without going through the selection process again.
This modular approach means you do not need to write a separate “Chain-of-Thought SQL Analyst” prompt and a “Few-Shot SQL Analyst” prompt. You mix and match the components you need for the task at hand. If you are interested in the reasoning techniques behind this approach, my article on agentic AI architectures covers how different prompting strategies fit into larger automation frameworks.
The prompt file format
Every prompt in the library follows the same format, which makes it straightforward to add your own. The file starts with YAML frontmatter:
---
title: Chain-of-Thought Prompting
category: frameworks
tags: [reasoning, step-by-step, problem-solving]
difficulty: intermediate
models: [claude, gpt-4, gemini, llama, mistral]
---
The difficulty field accepts beginner, intermediate, or advanced. The models field lists which LLMs the technique works best with. The tags field is what the search command matches against.
The body of the file follows a consistent structure: an overview paragraph, then “When to Use”, “The Technique”, “Template” (with {{placeholders}}), “Examples”, “Tips”, and “Common Mistakes”. The CLI and viewer both parse and display these sections.
Adding a new prompt is as simple as dropping a new .md file into the appropriate prompts/ subdirectory. The CLI picks it up automatically on the next run — no registration step, no build process.
What this is not
We want to be clear about the scope of this tool, because it is easy to imagine features it does not have.
The library does not connect to any AI model. It is purely a storage and retrieval system for prompt templates. You still need to copy the output and paste it into Claude, ChatGPT, Gemini, or whichever model you are using. There is no API integration, no automatic prompt execution, and no response handling.
👉 That’s not true anymore! With release 2.2.0 we added: AI Playground (send prompts directly to AI models and see responses in real time), API Settings panel (securely store API keys for OpenAI, Anthropic, and Google in your browser’s localStorage), Prompt Linter (analyze any prompt against 14 quality rules covering role definition, task clarity, output format, constraints, structure, examples, audience, tone, and more), Prompt Optimizer (automatically restructure prompts with best practices) and Smart Recommender (describe what you need in plain English and get personalized prompt suggestions from the library).
The search is also intentionally simple. It does keyword matching with a basic scoring system, not semantic search or vector similarity. If you search for “email” and the template uses the word “outreach” instead, it will not surface it. This is a deliberate trade-off to keep the tool dependency-free and fast.
Finally, the 52 prompts are a starting point, not an exhaustive collection. They reflect the use cases that come up most in our work. Your team will likely find gaps and want to add your own templates. The format is designed to make that easy.
Getting started
The quickest way to try it:
git clone https://github.com/diShine-digital-agency/ai-prompt-library
cd ai-prompt-library
node bin/prompt-lib.js list
Or install globally:
npm install -g @dishine/prompt-library
prompt-lib list
You need Node.js 18 or later. That is the only requirement.
If you want to start with the visual browser without touching the terminal, just open viewer.html directly in your browser. All 52 prompts are embedded in the file, so it works offline with no server needed.
The library is MIT licensed. Use it, fork it, extend it.