Create a library of reusable prompts

🎯 Why a prompt library?

The problem

Most professionals who use Claude or other LLMs daily share the same chaotic workflow:

1. Write a prompt in the chat
2. Get a good result
3. Forget the exact prompt
4. Start over from scratch next time
5. Get a different result (often worse)

It's like a developer writing code without ever saving it. Absurd, right?

Tangible benefits

Estimated ROI: for a team of 5 people using AI daily, a well-managed library saves 10-15h/week (according to a 2024 McKinsey study on AI productivity in the workplace).

📁 Library Architecture

Recommended File Structure

Organize your prompts in a clear tree structure: a root folder containing a README file for global documentation, a _templates folder for reusable base templates (articles, emails, analyses), followed by thematic subfolders such as marketing, content, code, analysis, and system-prompts. Each thematic folder can itself contain subfolders (for example, social-media with files for LinkedIn or Twitter).

Standard Format of a Prompt File

Each prompt in your library should follow a standard format structured into several sections: metadata (ID, version, category, author, date, tested model, quality score, tags), variables (in the form of a table with the name, description, and a concrete example), the system prompt itself, the user prompt with constraints, an expected output example, and the version history. For example, for a SaaS B2B prospecting email, you would define variables like "Jean Dupont" for the prospect's name, "TechCorp" for the company, "CTO" for the job title, "Infrastructure costs" for the identified problem, "CloudOptim" for your product, and "Your LinkedIn post about..." for the personalized hook. The system prompt would define the role of a senior B2B sales rep, and the prompt would ask to write a concise email (max 5 lines), with a short subject line that sparks curiosity, no cliché phrases, and ending with a single simple question.

🔧 The variable system

Variable types

Variables make your prompts dynamic and reusable:

Variables in practice

Let's take an SEO blog article template: the prompt targets a senior web copywriter and uses variables such as "TechFlow" for the company, "Les 10 erreurs de pricing SaaS" for the topic, "pricing SaaS" as the primary keyword, a list of secondary keywords like "tarification" or "modèle freemium", "fondateurs de startups SaaS" for the audience, "1500" for the target length, and "conversationnel" for the tone. The prompt also includes conditional variables, for example, the activation of a 5-question FAQ section compatible with schema.org, as well as constraints on the meta title and meta description.

Automating replacement

Variable replacement can be automated via a templating tool that scans the template, substitutes each key with its value (joining lists with commas), and handles default values. With OpenClaw, this variable replacement can be integrated into complete workflows, chaining multiple prompt templates without manual intervention.

🏷️ Categorization system

By domain

Organize your prompts by function: marketing for acquisition and conversion, content for web copywriting and social media, code for development and review, analysis for data and market analysis, operations for processes and documentation, sales for prospecting and closing, hr for recruitment and onboarding, and system-prompts for agents and chatbots.

By complexity level

By target model

Some prompts work better on certain models. Indicate this in the metadata with a compatibility table: for example, a prompt might score 9/10 on Claude 3.5 Sonnet, 7/10 on GPT-4 Turbo (with a tendency to be too long), work partially on Llama 3 70B which drops constraints, and 8/10 on Mistral Large. Use OpenRouter to easily test your prompts on all these models.

📊 Versioning and iteration

Git for prompts

Treat your prompts like code using Git: initialize a repository, create your folder structure, then make an initial commit. For each modification, use conventional commit messages like feat(marketing): email-prospection v2.3 - ajout variable HOOK for a new prompt, fix(catégorie): description for a fix, perf(catégorie): description for a performance improvement, test(catégorie): description for adding tests, or docs(catégorie): description for documentation. Tag stable versions (for example v1.0 for 25 tested and validated prompts).

Versioning workflow

The lifecycle of a prompt goes through four stages: first a draft tested with 10 different cases, then a beta version validated by a peer, next a stable version tagged v1.0, and finally continuous iterations based on real usage feedback. To structure this approach from the start, check out Le guide ultime du prompt engineering en 2025.

Effective iteration also relies on good prompting techniques. When a prompt doesn't deliver the expected results, methods like Chain-of-Thought or Few-Shot allow you to refine the output without starting from scratch — see Chain-of-Thought, Few-Shot, Tree-of-Thought : les techniques qui marchent to apply them concretely to your templates.

Prompt A/B Testing

To compare two versions of a prompt, set up an A/B testing system that randomly assigns a variant (A or B), records each result with a score and timestamped notes, then calculates the average of each variant to declare a winner. This allows you to iterate on your prompts in a data-driven way rather than by intuition.

🔗 Prompt composition and inheritance

Base templates (inheritance)

Create base templates that specific prompts inherit from. For example, a parent template _templates/base-article.md defines the general style (short sentences of 20 words max, active voice, concrete examples, no jargon) and the standard structure (optimized title, introduction with hook and promise, body in H2/H3, conclusion with CTA). A child prompt like content/blog-tech.md inherits from this template and adds specific instructions: include code blocks, add a "In summary" section with 3 bullet points, target a junior-mid developer, and link to the official documentation.

Composition (combining prompts)

You can also compose prompts by chaining several steps. For example, a complete monthly report combines four sequential prompts: an analysis of the month's data, an identification of trends based on the results of step 1, recommendations leveraging the insights from step 2 with a budget constraint, and finally a report drafting assembling the three previous results.

🤖 Automation with OpenClaw

OpenClaw is the ideal tool for operationalizing your prompt library:

Automated workflow

OpenClaw allows you to define multi-step workflows. For example, for a weekly tech industry newsletter: a first "Topic gathering" step uses a prompt template to generate 5 topics, a second "Writing" step takes these topics as input to write an 800-word body in a conversational tone, and a third "Optimization" step applies a copywriting optimization template focused on click-through rate.

Dynamic variables

OpenClaw can automatically inject variables from databases, external APIs, configuration files, or the conversation context, which virtually eliminates any manual input in your recurring workflows.

📈 Metrics and continuous improvement

What to measure

Tracking dashboard

To track these metrics, implement a simple logging system that records each prompt use in a structured file: prompt identifier, score given, model used, indication of manual modification, free-form notes, and timestamp. An aggregation function can then compile this data to provide, for each prompt, the total number of uses, the average score, the modification rate, and the list of tested models.

Improvement cycle

Each month, follow this process: identify the 5 most used prompts, check their average score, plan an iteration for those below 8/10, document why those above 9/10 work so well, archive prompts unused for 3 months, and share the best ones with the team.

🏢 Team management

Team conventions

Formalize your contribution rules in a CONTRIBUTING.md file: before creating a prompt, check that it does not already exist by searching tags, identify whether it can inherit from an existing template, and discuss the need with the team for any new domain. The format is mandatory (standard template with all metadata filled in, at least 2 variable examples and 1 expected output example). The validation process goes through a Git branch, testing on at least 5 different cases, a review by a colleague, and a merge only after approval with an average score of at least 7/10. For naming: files in kebab-case, variables in UPPER_SNAKE_CASE, and identifiers in the form CATEGORY-NNN (MKT-001, CTN-015).

Sharing and onboarding

Host your library on a private Git repository. For hosting your internal tools, Hostinger offers reliable and affordable server solutions. The repository's README must include a 5-step quick start (clone, browse categories, choose a prompt, fill in the variables, copy to the LLM) as well as library statistics: number of validated prompts, system prompts, contributors, and average score.

❌ Common mistakes

• Not standardizing the format: without a common template, each team member invents their own structure, making the library unusable for others.
• Ignoring versioning: modifying a prompt without keeping the history prevents rolling back when an iteration degrades the results.
• Too many variables: a prompt with 15 variables takes more time to fill out than to write from scratch. Limit yourself to 5-7 variables maximum per prompt.
• Not documenting output examples: without a concrete example, two people can interpret the same prompt in very different ways.
• Forgetting model compatibility: a prompt optimized for Claude can give poor results on GPT-4. Always indicate the target model in the metadata.

🛠️ Recommended tools

• Claude: for generation quality, particularly suited for structured prompts and long content.
• OpenRouter: to test your prompts on many models (GPT-4, Claude, Llama, Mistral) from a single interface.
• OpenClaw: to automate your prompt workflows with chaining, dynamic variables, and API integrations.
• Hostinger: to host your internal tools, tracking dashboards, and private Git repositories at a lower cost.

✅ The Essentials

• A structured prompt library saves a 5-person team 10-15 hours per week.
• Each prompt must follow a standard format: metadata, variables, system prompt, constraints, output example, and version history. To master the basics of this writing, Le guide ultime du prompt engineering en 2025 is a reference.
• The variable system (text, choice, list, boolean) makes your prompts reusable without rewriting.
• Use Git to version your prompts like code, with clear commit conventions.
• The complexity level (simple, intermediate, advanced) helps each team member find the right prompt.
• Inheritance and composition avoid duplication between similar prompts.
• A monthly improvement cycle (identify, score, iterate, archive) maintains the quality of the library over time.

🎯 Startup checklist

To create your library today:

1. Create the folder structure (15 min)
2. Initialize Git and make the first commit (5 min)
3. Migrate your 5 best existing prompts into the standard format (30 min). If some are not giving satisfactory results, Prompt debugging : quand l'IA ne comprend pas ce que vous voulez will help you fix them before integrating them.
4. Create 2-3 basic templates for your frequent use cases (30 min). For agents and chatbots, rely on the principles detailed in System prompts : l art de cadrer son IA.
5. Define the naming and variable conventions (15 min)
6. Share with your team and collect their prompts (ongoing)

The prompt library is an investment that pays off from the very first week. Every saved prompt is time gained forever. With the right tools — Claude for quality, OpenRouter for flexibility, OpenClaw for automation — you have everything you need to move from artisanal prompting to industrial prompting.
```