Languages & Platforms

Ruby • PHP • JavaScript • Swift • iOS/macOS

Gamer

Rubyist

Preferred AI - Claude

Languages & Platforms

Ruby • PHP • JavaScript • Swift • iOS/macOS

Gamer

Rubyist

Preferred AI - Claude

Languages & Platforms

Ruby • PHP • JavaScript • Swift • iOS/macOS

Gamer

Rubyist

Preferred AI - Claude

Javascript Bingo�

Has run a marathon

Knows the Konami code

Can name 5 DC sports teams

Been on TV

Pushed code to production today

​

Role

Who the AI should act as

"You are a senior software engineer"

Task

What should the AI do

"Create a database table schema..."

Tone

How should the output sound

"Be clear and concise"

Topic

What is the content about

"Used to create product recommendations"

Details

Specific requirements

"Our database is Postgres 18.1"

Format

Output structure

"Show schema in a table..."

The output is wildly inconsistent

Sometimes thoughtful and idiomatic. Other times technically correct but awkward, unmaintainable, or misaligned with intent.

This inconsistency isn't random

It's driven by how we frame problems and communicate expectations to the model.

You already have this skill

Clear communication: Writing expressive code, crafting commit messages, giving constructive feedback in code review.

Not Being Concise

Don't ramble on. Just include the information the AI needs to know to generate the output.

It doesn't have feelings... yet. 🤖

Missing Context

Asking to fix a bug, without including details such as possible causes, where in the application it is occurring, and potential relevant code.

Passive Voice

Being passive gives the AI room to meander. Active voice gives clearer direction to the end goal.

D

Define your goal

Be as specific as possible about what you want.

E

Explain the context

Languages, tools, purpose — narrow the output.

E

Establish a persona

Who should the AI act as?

P

Provide examples

Include similar content or code to guide output.

E

Enable questions

Let the AI ask for clarification first.

R

Refine and iterate

Tweak and expand based on the response.

★☆☆☆☆

Add user registration.

★★☆☆☆

Add user registration to a Ruby application, require an email, display name, and password.

★★★☆☆

...include a registration and sign in view. Store credentials in the database.

★★★★☆

...use the Devise gem. Ensure views are generated. Include tests for registration and sign in.

★★★★★

As a Senior Rails Engineer... use Devise... include tests... ask for clarification if needed.

But is the 5-star version actually the best approach? What's wrong with it?

At some point, I’m just rambling.

So what can be done to make this prompt better?

1

I want you to, as a Senior Ruby on Rails Engineer, add user registration to this application. The user will need to log in to access some areas. I will provide follow up prompts with more details. Stop and ask for clarification if you have any questions.

2

Use the Devise gem to add user registration. When a new user registers require an email, display name, and password.

3

Ensure the registration and sign in views are properly generated and match the theming of the rest of the application.

4

Include tests to verify users are able to register and sign in, including tests that all fields are required.

AGENTS.md

A markdown file with specific instructions for your agent to follow. Use them both globally and at the project level.

AI Skills

AI skills are reusable instruction sets that teach an agent how to perform a specific task well. They act as pre-built expertise the agent can reference on demand, so you don't have to re-explain the same process every time.

​

Think of them as playbooks: you write the best approach once, and the agent follows it consistently across projects.

​

Global

~/.config/agents/AGENTS.md

Instructions that always apply, regardless of project.

Project

/path/to/project/AGENTS.md

Instructions specific to a single project.

Guardrails

What is the agent not allowed to do? What files can it never read, modify, or delete?

Testing Expectations

What types of tests should be written? How detailed should test cases be?

Communication Preferences

When should the agent stop and ask? How verbose should responses and explanations be?

Code Conventions

Naming conventions, preferred patterns, and specific anti-patterns to avoid.

Project Context

What the project does at a high level. Tech stack: language, database, frameworks, dependencies.

Directory Structure

How the project is laid out. Where do assets, tests, and configs live?

Workflow

How to build, test, and lint the application. Git conventions.

Domain Knowledge

What do common acronyms and project-specific terms mean? Known edge cases.

Readability

Specify naming conventions, comment expectations, and how self-documenting you want the code to be.

Performance

"This runs on every request" or "This processes 10k+ records" changes the approach entirely.

Maintainability

Ask for code that follows existing patterns. Reference specific files or modules as examples.

Idiomacy

"Write this the way a senior Ruby developer would" produces very different output than "write this in Ruby."

Overconfidence

The AI will present hallucinated methods, gems, or APIs with complete certainty. If you don't recognize it, verify it.

Hallucinated Architecture

Ask for too much at once and the AI will invent abstractions, services, or patterns your project doesn't use.

Style Drift

Over a long conversation the AI loses context and drifts from your conventions. Watch for it.

Plausible but Wrong

The code runs, the tests pass, but the logic is subtly incorrect. AI is great at looking right.

Read it first

Don't copy-paste blindly. Read every line like you're reviewing a junior developer's PR.

Run the tests

Do the tests actually test meaningful behavior, or just assert that the code does what it does?

Check for hallucinations

Do the methods, gems, and APIs it referenced actually exist?

Look for shortcuts

Did it skip error handling, validation, or edge cases you'd flag in review?

Does it match the codebase?

Correct code in the wrong style is still a problem.

Debugging Flaky Specs

Paste the test, the failure output, and the relevant code. Ask the AI to identify the race condition or timing issue — not just fix the test.

Generating Boilerplate

Migrations, serializers, form objects, API clients. Give it one example from your project and ask it to follow the pattern.

Understanding Unfamiliar Code

Paste a method or module and ask "What does this do, step by step?" Great for onboarding onto a new codebase.

Refactoring

"Refactor this to use X pattern" with a concrete before/after example produces better results than "make this cleaner."

AI is a tool, not a teammate. It doesn't understand your system, your users, or your deadlines.

The best results come from developers who already know what good code looks like and use AI to get there faster.

Prompt engineering is a skill. Like testing, debugging, or code review. It gets better with deliberate practice.

Questions?

Slides Available

https://jcurcioconsulting.com/posts/llmday-nyc-engineering-better-prompts-for-ai-assisted-development

Actually, here is a QR code!