Catalog
concept#AI#Software Engineering#Architecture#Governance

Prompt Engineering

Methods for formulating and optimizing inputs to AI models to produce precise and consistent outputs.

Prompt engineering comprises principles and techniques for crafting inputs to large AI models to elicit desired answers, formats, and behaviors.
Emerging
Medium

Classification

  • Medium
  • Technical
  • Design
  • Intermediate

Technical context

OpenAI or other model APIsOrchestration tools (e.g., LangChain, LlamaIndex)Production channels (chat platforms, CMS, automation pipelines)

Principles & goals

Define goals clearly: Establish precise success criteria before designing prompts.Iterative testing: Continuously measure and refine prompts.Explicit context design: Structure and provide relevant context.
Build
Domain, Team

Use cases & scenarios

Compromises

  • Hallucinations or false facts resulting from imprecise prompts.
  • Leakage of sensitive information through uncontrolled context sharing.
  • Excessive automation without human oversight leads to quality issues.
  • Provide explicit examples (few-shot) to ensure desired format.
  • Separate system and user prompts and define responsibilities.
  • Integrate automated and regression tests into CI.

I/O & resources

  • Access to target model and API keys
  • Domain knowledge and example responses
  • Evaluation dataset and quality criteria
  • Prompt templates and variant catalog
  • Validation and test suites
  • Monitoring metrics and dashboards

Description

Prompt engineering comprises principles and techniques for crafting inputs to large AI models to elicit desired answers, formats, and behaviors. It combines linguistic precision, iterative testing, and context design. Goals are reliability, robustness, and efficient use of models across technical and product-focused applications.

✔

  • Fast prototypes and content with low development effort.
  • Improved output quality through targeted phrasing.
  • Reusable prompt patterns and templates.

✖

  • Dependence on model capabilities and versions.
  • Prompts are often sensitive to small changes.
  • Non-deterministic responses can hinder reproducibility.

  • Response accuracy

    Measure of percent correct or expected responses against a test set.

  • Throughput (requests per second)

    Number of processed prompt requests per time unit under production conditions.

  • Token cost per result

    Average cost in tokens or API fees to produce a valid output.

Support FAQ bot

A chatbot uses system prompts and example dialogues to deliver consistent answers for customer queries.

Automated content production

Marketing team generates varied product descriptions using prompt-based templates and evaluation metrics.

Invoice data extraction

Pipeline extracts structured fields from invoices via targeted extraction prompts and validation rules.

1

Set goals and success criteria; prioritize requirements.

2

Secure model access and run initial prompt experiments.

3

Structure prompts, create templates, and define validation.

4

Set up automated tests and monitoring; iterate improvements.

⚠️ Technical debt & bottlenecks

  • Unversioned prompt templates without change tracking.
  • Missing monitoring for quality regressions after model changes.
  • Hard-coded prompt strings scattered across services.
model-limitsdata-qualityprompt-variability
  • Using prompts that embed confidential data unencrypted.
  • Prompts intended to bypass safety rules or moderation.
  • Automated generation without quality checks in production.
  • False-positive evaluations due to too-small test sets.
  • Unintended overfitting to example responses.
  • Ignoring model updates and keeping outdated prompts.
Prompt design and linguistic precisionTesting, metrics and evaluation skillsBasic understanding of model architecture and cost models
Model capabilities (context size, temperature control)Latency and cost requirementsPrivacy and compliance requirements
  • • Model token and context size limits
  • • API rate limits and cost budgets
  • • Legal constraints on data usage and storage