# ai-engineering-skills

A Claude Code plugin marketplace with one plugin: **prompting**, whose prompt-engineering skill acts as a conversational consultant for designing, optimizing, and troubleshooting prompts — built on research-backed techniques, structural best practices, and a library of ready-to-adapt patterns.

## Installation

```
claude plugin marketplace add benjaminjackson/ai-engineering-skills
claude plugin install prompting@ai-engineering-skills
```

## prompting

One skill, **prompt-engineering**, backed by three reference files it loads as the conversation demands: a techniques guide, a best-practices guide, and a pattern library.

### prompt-engineering

A conversational consultant for prompt work. It doesn't hand you a template and walk away — it asks what you're actually trying to accomplish, sizes up the task, recommends techniques with the reasoning attached, and builds the prompt with you piece by piece. The point is to teach the methodology, not just deliver a prompt: every recommendation comes with a why, and cost (token budget, latency, repeatability) is weighed at every step.

#### How it works

The skill runs a four-phase workflow:

1. **Intake** — a few targeted questions to pin down the desired outcome, input format, constraints, target model, and what success looks like.
2. **Analysis** — classifies the task's complexity: simple tasks that zero-shot handles, medium ones that want few-shot examples or chain-of-thought, and high-stakes or exploratory ones that justify heavier techniques.
3. **Recommendation** — picks specific techniques via a decision framework (zero-shot for well-defined tasks, few-shot for format fidelity, chain-of-thought for multi-step reasoning, self-consistency when accuracy is critical, tree-of-thought for strategic exploration, least-to-most for multi-stage problems, self-refine for iterative improvement) and explains the trade-offs rather than just naming a technique.
4. **Construction** — builds the prompt component by component: role, instructions, delimiters and structure, examples, output format, constraints — explaining each choice, and ending with the complete prompt ready to copy plus advice on how to test and iterate.

It also has fast paths for common cases — a quick build for simple tasks, a diagnostic pass for "my prompt isn't working," an expert path for complex reasoning tasks, and a format path for "I need consistent output."

#### The references

- **techniques-guide.md** — deep dives on 10+ techniques (zero-shot, few-shot, chain-of-thought, self-consistency, tree-of-thought, generated knowledge, least-to-most, self-refine, directional-stimulus, role prompting) with decision criteria, pitfalls, and cost-benefit analysis.
- **best-practices.md** — core principles (clarity, positive guidance, separating instructions from content), role and system-message design, context organization, a six-step optimization process, common anti-patterns, and a quality checklist.
- **pattern-library.md** — ready-to-customize templates: classification, extraction, transformation, few-shot format learning, structured output (JSON, tables, hierarchies), multi-step workflows, and analysis patterns.

#### Usage

- **Explicitly:** ask directly — "Help me create a prompt for extracting structured data from customer emails," or "This prompt works but burns too many tokens, help me optimize it."
- **Automatically:** on phrases like "how should I prompt for X?," requests to improve or troubleshoot an existing prompt, or questions about prompting techniques ("when should I use chain-of-thought vs tree-of-thought?").

## Author

Benjamin Jackson ([@benjaminjackson](https://github.com/benjaminjackson))