50 Prompt Engineering Jokes for a Job That Did Not Exist in 2022

Prompt engineer: $340,000 base. $0 in actual training.

"Have you tried few-shot examples?" The universal solution to every problem.

Step 1 of every prompt engineering guide: "Be specific." Step 2: "Be more specific."

The system prompt is 4,000 tokens. The model still ignores the third bullet point.

"Let's add a chain-of-thought." The response is now 12 paragraphs.

Prompt engineer: Person who got the model to stop using emojis.

Asked the model to be concise. Added: "This is very important." Response: Three paragraphs.

"We've moved to temperature 0." The hallucinations got more confident.

Prompt engineering is mostly: Writing in all caps the third time.

The new prompt fixed bug #1. The new prompt broke bug #2 and bug #3.

"Treat this as a senior engineer." The model: Became more apologetic.

Spent two hours on the prompt. The user pasted "hi" and broke everything.

"You are a helpful, harmless, honest assistant." Four tokens of personality. The rest is the user's problem.

Prompt engineer interview: "Can you optimize this prompt?" Answer: Delete half of it. Result: Worked better.

"This is a critical task. Do not make mistakes." The model is now nervous and worse.

The team A/B tested two prompts. Neither prompt worked. They shipped both.

"Think step by step." The most powerful three words in software engineering, depending on the day.

Prompt engineering is just regex with extra steps and worse error messages.

Senior prompt engineer: Knows when to give up and use the API directly.

"We've prompt-tuned this model on internal data." The internal data: Four examples in the system message.

The prompt works on GPT-4. The prompt fails on GPT-4o. The prompt fails differently on Claude. The prompt does not exist for Gemini.

Half of prompt engineering is writing the prompt. The other half is explaining to PMs why the prompt did not solve the business problem.

Added "please" to the prompt. The response quality went up. Nobody can explain why.

The prompt is 1,800 tokens. The response is 1,800 tokens. The user's question was 6 tokens.

"Roleplay as a 10x developer." The model: Apologized in advance.

Prompt engineering meeting: Five people debating whether to add a comma.

"This prompt is final." The model released a new version on Tuesday.

Best prompt I ever wrote: The one I deleted to start over.

Job posting: "Prompt Engineer, 3+ years of experience." The field is 18 months old.

The prompt engineering candidate brought a portfolio. The portfolio: Screenshots.

Asked the model to ignore previous instructions. The model did. The model also ignored the new instructions.

The prompt engineer added an example. The example became the new prompt.

"Use this format: JSON." The model used JSON. The JSON is wrapped in markdown.

Step 1: Write a prompt. Step 2: Watch the model do the opposite. Step 3: Add the word IMPORTANT. Step 4: Same result.

The team has 14 versions of the same prompt. Version 7 is in production. Nobody remembers why.

"Be terse." Response: Is terse. Follow-up question. Response: No longer terse.

The prompt engineering wiki has 240 pages. 239 are unused. The one that works is page 1, written six months ago, by an intern.

"You are an expert." The model: Still apologizes about being an AI in the second sentence.

Prompt engineering best practice #1: Write less prompt. Nobody does this.

The prompt engineer wrote 800 tokens of instructions. The model followed sentence 4 and ignored the rest.

Asked the model for a one-line answer. Got a 1,500-word Notion page with a table of contents and three callout blocks.

The team prompt library is a Google Doc. It has 91 comments. Every comment is "is this still the one we use"

The contested word in "Prompt Engineering" is "Engineering." The contested word in "Prompt Engineer" is also "Engineer." There is no third word.

Model bake-off results: Claude won at writing. GPT won at math. Gemini won at the benchmark Google made up last week.

Built an eval set of 200 examples. The examples were generated by the model being evaluated.

"We are fine-tuning the model." They are editing the system prompt.

Every problem in the planning doc was solved by RAG. Every problem in the postmortem was caused by RAG.

"We need a vector database." Three weeks of plumbing later, the index serves one query type, and grep would have answered it in 4 milliseconds.

The system prompt now includes: "The user is in a good mood today." The response quality went up. Nobody can explain why.

The job title used to be Prompt Engineer. This quarter it is Context Engineer. Next quarter it will be something else. The work is the same.