New prompts and trending repos: week 46
The short answer
This week's roundup covers new research-agent and browser-automation prompts, a revised RAG evaluation guide, and trending repositories in agent sandboxing, local inference and document parsing. All links go to the original prompts and GitHub projects.
This week at AIPrompt123
We published six new prompts, revised the RAG evaluation guide and added four repositories to the GitHub directory. Here is what is worth your time.
New prompts
1. Write a system prompt for a research analyst. The important addition is the citation rule: every factual claim needs a source, primary sources beat secondary ones, and the model must flag uncertainty instead of smoothing it over.
2. Design a browser automation agent. Browser agents are powerful and easy to misuse. This prompt requires a domain allowlist, confirmation before submissions and a screenshot log.
3. Build an evaluation set for an agent. Twenty test cases across happy path, ambiguity, tool failure and adversarial input. If you only read one prompt this week, make it this one.
4. Write a customer service macro library. Twelve templates for the situations that consume the most support time.
5. Create a Hindi YouTube script. Written for natural Hindi with English technical terms where that is how people actually speak.
6. Design a content moderation agent. A policy-first design with severity levels, appeal handling and a human review threshold.
Updated guides
The RAG guide for product teams now includes a minimal evaluation plan with five metrics and a note on citation accuracy. The system prompt guide was revised with a section on instruction drift and multi-turn testing.
Trending repositories
browser-use continues to gain stars as teams experiment with browser agents. The appeal is obvious: a model that can operate a real website. The risk is equally obvious. Read the prompt on confirmation steps before you point it at anything with a submit button.
vLLM is trending again after a throughput-focused release. If you are serving open models on GPUs, it remains the default to benchmark against.
ComfyUI keeps growing as image and video workflows become more complex. The node graph is still the best way to make a generation pipeline reproducible.
RAGFlow is attracting attention for document parsing quality. If your PDFs have tables and multi-column layouts, parsing is the part that decides whether retrieval works at all.
One thing we learned
We reviewed 40 user-submitted prompts this week. The most common problem was not the instruction; it was the missing output format. A prompt that says "write a good summary" produces something different every time. A prompt that specifies length, sections and what to exclude produces something you can use.
If you are submitting a prompt, include the output contract. It is the difference between a demo and a workflow.
From the library
- The best ChatGPT prompts for students
- Midjourney prompt formula
- Top GitHub repositories for AI developers
Subscribe
Get the next roundup by email. One useful prompt, one repository worth starring, and the week's AI news without the hype. Join the newsletter.
Copy-ready prompts featured in this guide
Copy any prompt below or open its full page with interactive variable fields and platform tabs.
Site Admin
Author & Platform Administrator
Administrator and lead editor at AIPrompt123, curating high-performance AI prompts, SEO/AEO/GEO strategies, and developer repositories.
Questions about this guide
Short answers to what readers ask most about this topic.
Keep reading
Top GitHub repositories for AI developers in 2026
The frameworks, inference engines, agent tools and learning resources that are actually worth your time, with licence notes and what each is best for.
30 AI prompts for small business owners (with the tasks that matter)
Marketing, customer service, pricing and admin prompts for teams without a dedicated AI department, plus the guardrails to keep you out of trouble.
The best ChatGPT prompts for students (that actually help you learn)
Most 'study prompts' generate summaries that feel productive and teach you nothing. These twelve prompts are built around retrieval, feedback and self-testing.