AI Guild
Table of Contents
Datadog Dashboard Generation
Prompt for comprehensive metrics tracking
E2E Bug Fix with Browser
Claude fixing bugs using --chrome flag
AWS Cost Analysis
Full cost audit using AWS CLI
Requirements-First with Claude Code
Turning a design doc into a full project
This Presentation
Done end-to-end by Claude
PRESENTED BY
Jaume Puig
CTO @ Enginy
Datadog Dashboard
One prompt to generate comprehensive metrics
Can you please add as many metrics as possible to get a good picture
of what is happening with credits:
- How many times modals are opened
- Which modals are being opened
- When users click on various elements
- User interaction patterns
- etc.
Make it VERY VERY DETAILED.
And then give me JSON to create a dashboard in Datadog.What Claude does:
- Analyzes the codebase for all credit-related flows
- Identifies all modal components and user interactions
- Adds detailed tracking events throughout
- Generates ready-to-import Datadog JSON
Result:
- 50+ new tracking events added
- Complete visibility into credit usage
- Dashboard ready in minutes, not hours
E2E Bug Fix: Post Likes
Browser verification + Temporal workflow testing
Browser: Verify the Bug
Run with browser access
claude --chrome
Navigate to post page
Claude sees the like button isn't working
Verify the fix visually
Takes screenshot to confirm post is now liked
Temporal CLI: Test the Fix
Claude runs the workflow directly to test the backend fix:
# Claude executes the workflow to test
temporal workflow execute \
--type PostLikeWorkflow \
--task-queue main \
--input '{"postId": "123"}'Sees the workflow succeed, confirming the backend fix works.
Key insight: Claude verifies fixes end-to-end: visually in the browser AND by running the actual Temporal workflow. No manual testing needed.
AWS Cost Analysis
Full audit using AWS CLI
My prompts to Claude Code:
Use aws cli to see where we are spending money and how we could save money
Provide a pdf report with how we could save money
Also provide an .md file with all the steps that you did
AWS CLI Commands Claude Ran:
aws ce get-cost-and-usage --group-by SERVICE
aws ce get-savings-plans-purchase-recommendation
aws ce get-reservation-purchase-recommendation
aws ce get-cost-and-usage --filter SERVICE --group-by USAGE_TYPE
Key Findings:
Savings Plans
$116/mo
ElastiCache RI
$15/mo
IPv4 Cleanup
$30-50/mo
CloudWatch Opt
$50-100/mo
Output: PDF report + markdown docs with 8 analysis steps. Claude used pandoc + weasyprint for PDF generation.
Requirements-First with Claude Code
A detailed design doc became a full project
Key insight: The more time you spend describing requirements, the better Claude Code performs. A detailed spec is not overhead - it's your best investment.
The Design Document: 12 Comprehensive Sections
1. Overview
Core concepts, key decisions
2. Database Schema
6 Prisma models, enums, indexes
3. API Specification
20+ tRPC endpoints with I/O
4. Business Rules
State machines, constraints
5. Error Codes
Standardized error handling
6. Security
Auth, authorization, audit
7. Future Scope
Explicitly out-of-scope items
8. Generation System
Temporal workflows, retries
9. Frontend Design
Pages, components, state
10. File Structure
Directory layout, naming
11. Type Definitions
Full TypeScript types
12. Testing Strategy
Unit, integration, E2E
2,200+
Lines of specification
6
Database models
20+
API endpoints
5
Frontend pages
What "Detailed Requirements" Look Like
Examples from the design doc
Key Design Decisions Table
API Endpoint Specification
// Input:
{
aiVariableId: string;
models: string[]; // ["gpt-4o", "claude-sonnet"]
questionIds: Array<{
questionId: string;
required?: boolean;
order?: number;
}>;
hypothesis?: string;
}
// Validations:
// - AI variable must exist and belong to user's client
// - At least 1 model must be specified
// - All questions must exist and be activePrisma Schema with Indexes
model HumanLoopItem {
id String @id @default(cuid())
experimentId String
model String
generatedContent String
filledPrompt String
entitySnapshot Json
status ItemStatus @default(pending)
displayOrder Int // Randomized for blind eval
@@index([experimentId, status])
@@index([experimentId, displayOrder])
}Business Rules: State Machine
Experiment Status Transitions:
Each transition has explicit validations defined
Why this matters: Claude Code can implement exactly what you need when you specify exactly what you want. Ambiguity leads to rework.
The Workflow: From Spec to Working Code
My role vs Claude Code's role
Write Design Doc
Me: Drafted the full spec with Claude's help in regular chat. Iterated on edge cases, naming, and structure.
Implementation
Claude Code: Read the doc, created Prisma migrations, tRPC routers, Temporal workflows, React pages.
Review & Iterate
Together: I reviewed PRs, pointed out issues, Claude Code fixed them. Minimal back-and-forth.
What Claude Code Produced From the Spec:
🗂
Prisma Schema
6 models + migrations
🔌
tRPC Router
20+ procedures
⚡
Temporal Workflows
Generation + retries
🖥
React Pages
5 pages + components
📋
Types
Full TypeScript coverage
Takeaway: I spent ~2 hours on the design doc. Claude Code implemented the full feature in ~30 minutes of prompting. The ratio of thinking vs coding has completely flipped.
This Presentation
Done end-to-end by Claude
This entire presentation was created by Claude Code with a single prompt:
Create a new presentation using the other ones as templates.
The presentation should be for AI Guild.
I want to present these three things, each one in one slide:
- [Metrics prompt example]
- [E2E bug fix example]
- This presentation, done e2e by Claude🔍
Explored templates
Read existing presentations
📁
Created directory
Set up folder and assets
✍️
Wrote the HTML
Generated all slides
☁️
AWS + CI/CD
S3, CloudFront, GitHub Actions
My role: Prompting + reviewing each step. Claude did all the coding, file operations, AWS infra setup, and CI/CD pipeline.
Questions?
AI Guild