Prompt Engineering

Few shot prompting - give it a few examples in your prompt to follow for output
Decomposition - what are the sub problems you should solve first before you solve this problem?
Self-criticism - ask it to criticize a previous output as a way to improve its output or your prompt. “Check your response and reflect on the answer.”
Additional information/ additional context - more is better
Ensemble - use different approaches and different models, extract what is common across them
For some non-reasoning models, invoke “deep thought” by asking it to walk through the steps to its answer. This also still applies to reasoning models, sometimes. Ask it to “think hard”
Use XML tags for compartmentalizing parts of the prompt, or for organizing the prompt into sections
Include a “failed example” in the prompt to inform the model what not to do
Role-prompting - “you are aggressive-email-composer-52392”, works well for expressive writing tasks
Meta-prompting - feed in your system prompt to an LLM and ask it to improve it

Last modified: July 12, 2025