Hermes Agent Prompt Library — Curated AI Prompt Templates¶

What's Inside¶

Collection	File	Use For
Code Generation	`code-generation.md`	Writing, refactoring, debugging, and reviewing code
Content Creation	`content-creation.md`	Blog posts, social media, email, video scripts, SEO
Data Analysis	`data-analysis.md`	SQL queries, reporting, visualization, metrics
Business Operations	`business-operations.md`	Email, meetings, project planning, task management
Research	`research.md`	Competitive analysis, market sizing, tech evaluation
Creative	`creative.md`	Brainstorming, naming, design briefs, creative strategy

How to Customize Prompts¶

The Bracket Convention¶

Every prompt uses [BRACKETED TEXT] as a placeholder. Replace these with your specific details before sending the prompt to any model. For example:

Template: "Write a [WORD COUNT]-word blog post about [TOPIC] targeting [AUDIENCE]."

Filled: "Write a 1200-word blog post about Kubernetes cost optimization targeting DevOps engineers at Series B startups."

The more specific your replacements, the better your output. Vague inputs produce generic results.

Adding Your Context¶

Most prompts work better when you add:

Examples of your style. For emails, paste 2-3 emails you've written that you like. For code, include a function from your codebase that shows your patterns.
Your data schema. For analysis prompts, include actual table names and column types.
Your constraints. Mention specific frameworks, brand guidelines, budget limits, or technical requirements.
Your definition of success. What does "good" look like for this specific output?

Combining Prompts¶

These prompts are designed to be chained. Common pipelines include:

Code pipeline: code generation → code review → test generation → documentation
Content pipeline: research prompt → outline → first draft → editing pass
Analysis pipeline: exploratory analysis → anomaly detection → report generation
Operations pipeline: task breakdown → sprint plan → daily standup → weekly recap

Model Selection Guide¶

Not all prompts work equally well with all models. Here's a general guide:

Code Generation and Debugging¶

Best with: Claude 3.5 Sonnet, GPT-4o, DeepSeek V3, Gemini 2.5 Pro Why: These models excel at structured reasoning, syntax accuracy, and following detailed technical specifications. For simple boilerplate, smaller local models like CodeLlama or DeepSeek Coder can handle it efficiently.

Content Creation¶

Best with: Claude 3.5 Sonnet, GPT-4o Why: Strong prose generation, tone control, and brand voice consistency. For high-volume social content, smaller models fine-tuned on your brand voice can be more cost-effective.

Data Analysis¶

Best with: Claude 3.5 Sonnet (long context for schemas), GPT-4o Why: SQL generation accuracy, ability to reason about complex schemas, and thorough edge-case handling. Provide actual schema definitions — don't summarize them.

Business Operations¶

Best with: Claude 3.5 Sonnet, GPT-4o, Gemini Flash Why: Email and meeting tasks benefit from nuanced understanding of professional context and relationship dynamics. For quick drafts, smaller/faster models work well; for sensitive communications, use a frontier model.

Research¶

Best with: Claude 3.5 Sonnet (200K context for document analysis), GPT-4o with web search Why: Research tasks often involve synthesizing large amounts of information. Long-context models handle document dumps better. Web-enabled models can pull in current data.

Creative Work¶

Best with: Claude 3.5 Sonnet, GPT-4, Gemini Why: Creative tasks benefit from models that take interesting "risks" in their outputs. Temperature settings matter — try 0.8-1.0 for brainstorming, 0.5-0.7 for refined creative output.

Prompt Chaining Strategies¶

Start with a broad prompt, evaluate the output, then follow up with targeted refinements: 1. "Write a blog post about [TOPIC]." → gets a first draft 2. "Make the introduction more personal, add a customer story." → improves specific sections 3. "Shorten everything by 20% and add subheadings." → tightens and structures

Parallel Generation¶

For creative tasks, run the same prompt 3-5 times (or ask for 3-5 variants in one prompt) and select the best. Models are non-deterministic — different runs produce different ideas.

Role-Based Chaining¶

Use different system prompts for different stages: 1. Researcher persona: gather facts and outline structure 2. Writer persona: draft the content 3. Editor persona: review, tighten, improve flow 4. Fact-checker persona: verify claims and data

Common Pitfalls and Fixes¶

Problem	Likely Cause	Fix
Output is too generic	Placeholders too vague	Add specific constraints, examples, audience details
Model ignores a key requirement	Buried in long prompt	Put critical requirements at the top or repeat them
Code doesn't run	Missing context about environment	Specify language version, framework, dependencies
Creative output is bland	Temperature too low	Use temperature 0.8-1.0; ask for "surprising" or "unconventional" ideas
Analysis is shallow	Didn't provide schema or data sample	Give the actual structure, not a description of it
Email feels robotic	No style examples given	Paste 2-3 examples of your own writing as reference

Contributing¶

These prompts are starting points. As you discover effective variations, adapt them to your workflow. The best prompt is the one that consistently produces useful output for your specific context — invest in refining it over time.