Reading Time: 5 minutes

Hey Prompt Lover,

Eighteen newsletters in. Still March. Still The Prompt Report.

Before we get into today's technique I need to tell you something about where we are in this series.

We started this in February. Seventeen newsletters covering fifteen techniques pulled directly from a 200-page research paper covering 1,565 studies on prompting and prompt engineering. Module by module. Technique by technique. Every prompt tested before it landed in your inbox.

We have covered a lot of ground together.

The five-component structure. Prompt sensitivity. Role prompting. Zero-shot techniques. Few-shot prompting and why example order swings accuracy by 40 points. Chain-of-Thought. Contrastive examples. Self-Consistency. Decomposition. Tree-of-Thought. Self-Refine. Chain-of-Verification. Meta-prompting. ProTeGi. Answer engineering.

Seventeen techniques. Seventeen working prompts. All of it from the research.

Here's where we are now.

We have three newsletters left in this series. This one and two more.

Three newsletters to cover everything that remains in The Prompt Report. Multilingual and multimodal prompting. Agents and RAG. Evaluation. Security. Sycophancy and bias. The benchmarking results. And the 47-step case study — the most honest piece of research writing I've read in years.

We're not going to rush through the remaining material. We're going to be selective. Everything that has direct practical value for your work makes it in. Everything that belongs in a research lab and nowhere else gets a sentence and moves on.

Today we start with the most counterintuitive finding in the entire paper.

Late last year a client asked me to help her build prompts for a content workflow that needed outputs in three languages. English, Spanish, and Portuguese. Her audience was across Latin America and she wanted the content to feel native, not translated.

My assumption going in was obvious. Write the prompts in the language you want the output in. You want Spanish output, write a Spanish prompt. That's just logical.

I was wrong. The research says otherwise and I tested it myself to be sure.

The finding from The Prompt Report is this: English prompts consistently outperform prompts written in the task language, even when the output needs to be in a completely different language.

Not sometimes.

Consistently.

Across multiple models and multiple non-English tasks.

I built the same prompt in English and in Spanish. Same role, same directive, same context, same examples. The English version produced better Spanish output than the Spanish version did.

Not better English output. Better Spanish output.

I've thought about why this is true more than almost any other finding in the paper. The explanation the researchers point to is straightforward once you hear it: training data is English-heavy.

The model's strongest reasoning capabilities were built on English text. When you prompt in English, you're working with the model at full capacity.

When you prompt in another language, you're working with a model that reasons slightly less precisely in that language regardless of how well it can produce output in it.

Reasoning and output are different. The model can reason in English and output in Spanish.

That combination outperforms reasoning and outputting in Spanish alone.

This finding matters beyond multilingual work.

It's a reminder that AI models are not symmetric. They don't have equal capability across every input configuration. The same model that produces excellent output from an English prompt may produce noticeably weaker output from a structurally identical prompt in another language.

Most people who work with non-English content never test this because the assumption is so logical it never gets questioned. Of course you write Spanish prompts for Spanish content. Why would you do anything else.

Because the output is better when you don't. That's why.

The research also found something worth knowing about translated prompts specifically. Human-translated prompts outperform machine-translated prompts. If you're going to translate a prompt into another language, do it yourself or hire a native speaker. Running your English prompt through a translation tool and using that as your non-English prompt produces weaker results than just keeping the prompt in English.

By the end of this issue, you'll have:

• The research finding on multilingual prompting and the specific technique that improves cross-language outputs

• The core multimodal techniques that transfer directly from everything we've covered in text prompting

• A working prompt template for multilingual tasks

• A clear picture of where Agents and RAG fit into your existing prompt workflows

Let's get started.

While many businesses are solving problems at lightspeed, their payroll systems seem to stay stuck in the past. Deel's free Payroll Toolkit shows you what's actually changing in payroll this year, which problems hit first, and how to fix them before they cost you. Because new compliance rules, AI automation, and multi-country remote teams are all colliding at once.

Check out the free Deel Payroll Toolkit today and get a step-by-step roadmap to modernize operations, reduce manual work, and build a payroll strategy that scales with confidence.

Download the Toolkit for free today

The research documents two techniques specifically for multilingual tasks. Both are worth knowing.

Translate First: Before prompting with non-English input, translate it to English. Process it in English. Then translate the output back if needed. Works best when the reasoning step is complex and the input language is one the model handles less precisely.

XLT — Cross-Lingual Thought: Combines three elements in one prompt. A role that establishes cross-lingual expertise. An instruction to think through the problem in English before producing the output in the target language. And a chain-of-thought step before the final answer.

The paper covers image, audio, video, and 3D prompting. Here's what's actually useful right now.

For image generation:

Two techniques that work immediately.

Negative prompting: instead of only describing what you want, explicitly describe what you don't want and downweight it.

"No text overlays.

No blurry backgrounds.

No stock photo lighting."

This is the same principle as Contrastive Chain-of-Thought from Newsletter 9. Showing the boundary between what you want and what you don't produces better results than describing the destination alone.

Paired-image prompting: show a before and after transformation, then show a new image and ask for the same transformation.

You're demonstrating the pattern rather than describing it. Few-shot prompting applied to images.

For image analysis:

Image-as-Text: generate a detailed text description of the image first, then include that description in your text-based prompt.

Useful when your AI tool handles text prompting better than multimodal input, or when you need to combine image content with complex reasoning instructions.

The core principle across all of it: every text prompting technique has a multimodal equivalent. Few-shot works with images. Chain-of-Thought works with images. Role prompting works with images.

The techniques transfer. Most people haven't tried them because they think multimodal prompting is a different skill. It isn't.

We're not going to spend three newsletters on agents. We're going to spend two paragraphs on what you need to know and one prompt on what you can use.

What agents are: AI systems that interact with external tools to complete tasks. Instead of relying on what the model already knows, agents can search the web, run code, read documents, use calculators, query databases. The model becomes a router that decides which tool to use and when.

What RAG is: Retrieval Augmented Generation. Instead of asking AI to remember information, you retrieve the relevant documents and insert them into the prompt. AI reads your documents and answers from them. Stops hallucination on knowledge-intensive tasks more effectively than any prompt instruction.

Why both matter: They solve the same underlying problem from different angles. AI is unreliable when it has to rely on training data for specific factual claims. Agents and RAG both give the model access to accurate information at the point of need rather than asking it to remember things it might have wrong.

The practical version most people can use today:

Copy and paste the relevant document, data, or context directly into your prompt before asking your question. That's RAG without any technical setup. You're retrieving and inserting manually. The principle is identical. The output reliability improvement is immediate.

▼ COPY THIS PROMPT — MANUAL RAG:

Modules 7 and 8 of The Prompt Report cover techniques that feel more advanced than everything that came before them. Multilingual optimization. Multimodal reasoning. Agents. Retrieval systems.

But every single one of them is built on the same foundation we laid in Newsletter 1.

The five components. The role. The context. The examples. The format. The directive.

Multilingual prompting is still a structured prompt. Multimodal prompting is still few-shot and Chain-of-Thought applied to a different input type. Manual RAG is still context in the prompt body, which is Newsletter 1.

The techniques get more sophisticated. The structure underneath them stays the same.

That's not a coincidence. It's the most important thing The Prompt Report confirms across 1,565 papers. The foundation matters more than the technique. Get the structure right and every advanced technique works. Skip the structure and nothing works consistently.

Next newsletter we cover four things that will make you permanently more careful about how you use AI outputs. Security and prompt injection — the documented ways AI tools get exploited and what actually prevents it. Sycophancy — the research finding that larger, smarter models are more likely to tell you what you want to hear even when you're wrong.

Bias — how your example choices shape outputs in ways you can't see. And the benchmarking results — what happened when the researchers tested six techniques head to head on 2,800 questions and which one actually won.

The final newsletter is the 47-step case study. One researcher. One prompt. Twenty hours of work. And the automated system that beat him in sixteen iterations.

The most honest piece of writing in the entire paper and the best summary of what prompting actually is that I've found anywhere.

Two newsletters left. They're worth staying for.

If you work with any non-English content, take your best performing English prompt and run it against the same prompt written in the target language. Same task. Same context. Two languages.

Compare the outputs.

If the English version produces better target-language output, you now have a finding from your own workflow that confirms what the research found across 1,565 papers.

If the target-language version wins, reply and tell me. I want to know.

Test the English versus native language prompt on something real and reply with what you found.

Or try the manual RAG prompt on a document-based question and tell me if the output reliability changed.

Two newsletters left. I'll be here for both of them.

— Prompt Guy

P.S. Eighteen newsletters in. Two more to go. The series wraps this month with the benchmarking results and the 47-step case study. If you've been reading since Newsletter 1, thank you. If you're just finding this, reply and I'll send you the full archive. Everything builds on everything else. Starting from the beginning is worth it.