Your guide through product prioritization and backlog management
Built for Product Owners, Product Managers, and anyone managing product backlogs
Atlas is a conversational AI assistant designed to help Product Owners and Product Managers reduce time spent on repetitive operational tasks. Rather than replacing product judgment, it automates the mechanical aspects of backlog management, allowing product professionals to focus on strategic decisions.
Works in English and Spanish - automatically detects your language and responds accordingly.
Built by: Ghiselle Butrón
Timeline: 1 week (November 2025)
Tech Stack: Poe.com, Claude-Sonnet-3.5, Prompt Engineering
Type: MVP concept validation / Learning project
Transform vague feature descriptions into well-structured user stories following best practices.
Example:
- Input: "Users need to save searches"
- Output: "As a registered user, I want to save my favorite searches, so that I can access them quickly without searching again"
Objectively rank features using the RICE framework with guided inputs:
- Reach: How many users will this impact per quarter?
- Impact: How much will it impact each user? (0.25=minimal to 3=massive)
- Confidence: How confident are we in our estimates? (percentage)
- Effort: How much work is required? (person-months)
Formula: RICE Score = (Reach × Impact × Confidence) / Effort
Generate clear, testable acceptance criteria for user stories using Given-When-Then format when appropriate.
Get relevant KPI suggestions tailored to your feature type, with explanations of why each metric matters.
Atlas automatically detects whether you're writing in English or Spanish and responds in your language, maintaining the same professional quality and structure in both languages.
Great question. For users skilled in prompt engineering, using Claude, ChatGPT, or any AI chat directly can achieve similar results.
Atlas isn't about revolutionary technology—it's about product design and accessibility.
1. Reduced Friction
- No need to craft prompts from scratch each time
- Pre-optimized prompts tested across 15+ real scenarios
- Consistent structure and format
- No "blank page syndrome"
2. Democratized Access
- Not everyone managing product is a prompt engineering expert
- Makes AI capabilities discoverable through clear options
- Guides users through the process step-by-step
- Works the same way every time, regardless of AI experience
- Available in English and Spanish for broader accessibility
3. Standardization
- Teams get consistent user story formats
- RICE calculations follow the same methodology
- Reduces variability in output quality—critical when multiple PMs/POs are writing stories
- Everyone uses the same framework
4. Focused Experience
- Purpose-built for product management tasks vs general-purpose chat
- Relevant examples and guidance built-in
- Less cognitive load to get started
- No distractions from unrelated capabilities
Think of it as:
General AI chats (Claude, ChatGPT, Gemini) are like having a full kitchen with every tool and ingredient imaginable.
Atlas is an optimized recipe card for specific dishes you cook regularly—you still use the same ingredients, but the instructions are pre-written, tested, and ready to follow.
| Aspect | General AI Chat | Atlas |
|---|---|---|
| Getting Started | Blank prompt - what do I ask? | 4 clear options to choose from |
| Prompt Quality | Depends on your skill | Pre-optimized through 15+ tests |
| Consistency | Varies by how you ask | Same structure every time |
| Learning Curve | Need prompt engineering skills | Guided experience for everyone |
| Context | General purpose assistant | Purpose-built for PM tasks |
| Language | May mix languages or miss nuances | Detects and responds in your language |
-
Research (2 hours)
- Interviewed 3 Product Owners about repetitive tasks
- Identified pain points: user story writing, prioritization, criteria definition
-
Design (3 hours)
- Mapped workflows for each task
- Defined 4 core functionalities based on frequency and impact
- Prioritized using RICE (meta!)
-
Prompt Engineering (5 hours)
- Crafted structured prompts for each feature
- Iteratively refined based on output quality
- Tested edge cases and ambiguous inputs
- Added multi-language detection and response capability
-
Model Selection (1 hour)
- Compared GPT-4o-Mini vs Claude-Sonnet-3.5
- Evaluated on: consistency, formatting, instruction-following
- Result: Claude superior for structured PM outputs
-
Testing (3 hours)
- Validated with 15+ real-world scenarios
- Adjusted prompts based on failure modes
- Optimized for clarity and consistency
- Tested language detection accuracy
-
Documentation (2 hours)
- Wrote comprehensive README with project overview, learnings, and roadmap
- Captured screenshots demonstrating the 4 core features
- Documented development process and technical decisions
- Created Spanish version of documentation
-
Platform: Poe.com (no-code conversational AI builder) was a strategic choice for this MVP phase, prioritizing speed and learning over infrastructure complexity.
- Why on Poe:
- Rapid prototyping: No-code platform enabled idea-to-prototype in days instead of weeks
- Focus on core value: Time spent on UX and prompt optimization, not API management or hosting setup
- Model flexibility: Native support for testing multiple AI models (Claude, GPT-4) without separate API integrations
- Lean validation: Perfect for testing the core hypothesis—"Is an AI assistant for PM tasks genuinely useful?"—before investing in custom development
- Why on Poe:
-
AI Model: Claude-Sonnet-3.5 (Anthropic)
- Why Claude over GPT:
- Superior instruction-following for structured outputs
- More consistent formatting (critical for user stories and criteria)
- Better at maintaining conversation context within session
- Natural, professional tone matching PM communication style
- Lower hallucination rate in factual content
- Excellent multi-language capabilities with consistent quality
- Tradeoff: Slightly slower response time vs GPT-4o-Mini, but quality justifies it
- Why Claude over GPT:
-
Core Technique: Advanced prompt engineering with structured outputs
-
Design Approach: Conversational UX principles
Model Selection Rationale:
After testing both GPT-4o-Mini and Claude-Sonnet-3.5 with identical prompts, Claude consistently produced:
- Better-formatted user stories (cleaner structure)
- More reliable RICE calculations (fewer math errors)
- More professional tone in outputs
- Better adherence to Given-When-Then format in acceptance criteria
- More natural language switching between English and Spanish
For a tool where consistency and precision matter, Claude was the clear choice.
Product thinking over technical complexity. The goal wasn't to write code, but to design a user experience that solves real problems through the right tool selection and thoughtful prompt design.
💡 Prompt structure is 90% of output quality
The way you frame requests, provide context, and structure expected outputs matters far more than raw model capability.
💡 Model selection is a product decision
Choosing Claude over GPT wasn't about "better AI"—it was about fit for purpose. Claude's strengths (structured outputs, consistency) aligned perfectly with PM tasks. This is product thinking applied to tech choices.
💡 Real scenarios ≠ theoretical scenarios
Edge cases and ambiguous inputs only emerge through actual testing. What works in theory often fails in practice.
💡 Examples dramatically improve consistency
Including concrete examples in prompts (few-shot learning) creates more predictable and useful outputs across both models.
💡 Language detection is a product feature
Adding automatic language detection wasn't just technical—it was about accessibility. Spanish-speaking PMs can now use Atlas naturally without switching mental contexts.
💡 Conversational UX is product design
How the bot asks clarifying questions, guides users, and structures information is as important as the quality of final outputs.
💡 Constraints drive prioritization
Platform limitations (like Poe's usage caps) forced me to focus on the highest-impact features first—a valuable PM exercise.
💡 Accessibility matters more than sophistication
A simple tool that anyone can use beats a powerful tool that requires expertise. Supporting multiple languages is part of true accessibility.
💡 Shipping beats perfecting
Better to have a functional MVP with documented limitations than an unfinished "perfect" solution.
💡 Product value ≠ technical complexity
The most valuable products often solve problems through design and UX, not advanced technology.
If this were deployed as a production tool, I would measure success through:
| Metric | Target | Why It Matters |
|---|---|---|
| Time saved per task | 60% reduction | Validates core value proposition |
| Adoption rate | 70% of POs using weekly | Indicates product-market fit |
| Output quality score | 4.5+ / 5.0 | Ensures generated content is actually useful |
| Weekly return rate | 80% active users | Measures stickiness and real value |
| Tasks completed | 500+/week | Shows sustained usage and utility |
| Language distribution | Track Spanish vs English usage | Validates multi-language value |
- ✅ Generates structurally correct user stories 95% of the time
- ✅ RICE calculations are mathematically accurate 100% of the time
- ✅ Acceptance criteria are testable and specific
- ✅ Users can complete tasks 3x faster than manual approach
- ✅ Language detection works accurately 100% of the time
If this were to evolve into a production product, next iterations would include:
- Jira/Linear sync - Direct backlog import and export
- Slack/Teams integration - Access Atlas from where teams work
- GitHub integration - Link stories to code
- Project memory - Maintain context across sessions
- Team learning - Adapt to company-specific terminology
- Historical analysis - Learn from past prioritization decisions
- Multi-user sessions - Collaborative prioritization workshops
- Template library - Share best practices across teams
- Commenting & feedback - Iterate on generated content
- Configurable frameworks - Beyond RICE (Kano, etc.)
- Multi-language support - Spanish, English ✅ IMPLEMENTED
- Additional languages - Portuguese, French, German
- Usage patterns - Understand bottlenecks
- Quality metrics - Track output effectiveness
- Usage limits: Poe free tier operates on a daily points system - users receive points that reset every 24 hours, allowing approximately 10-15 conversations per day depending on conversation length
- No persistence: Does not remember past conversations or projects
- No integration: Works standalone, doesn't sync with PM tools
- No project context: Doesn't know your product, users, or constraints
- Requires validation: All outputs need human review and refinement
- Generic advice: Can't access your specific data or metrics
- Limited to two languages: Currently supports English and Spanish only
For a real deployment, I would:
- Use Anthropic API directly (eliminates platform limits)
- Build database layer for project context and history
- Implement auth and team management
- Create integration layer for Jira, Slack
- Add feedback mechanisms to improve over time
- Expand language support based on user demand
This MVP proves the concept. Production would require a couple months of development.
✅ Problem Identification
Recognized that POs spend 30-40% of time on operational tasks that could be streamlined
✅ Solution Design
Designed focused solution addressing 4 high-impact use cases vs trying to solve everything
✅ Prioritization
Applied RICE framework to features themselves—chose quick wins over nice-to-haves
✅ User-Centric Thinking
Optimized for accessibility (non-experts) over sophistication (power users). Added language support for broader reach.
✅ Iterative Development
Rapid MVP → test → refine cycle over 1 week
✅ AI Application Knowledge
Understands where AI adds value and where it doesn't
✅ Prompt Engineering
Practical experience crafting effective prompts for consistent outputs across multiple languages
✅ Model Selection
Made informed technical decisions based on use case requirements, not hype
✅ Limitations Awareness
Clear-eyed about what this is (design project) vs what it isn't (technical innovation)
✅ Rapid Prototyping
Idea to functional demo in 1 week
✅ Professional Documentation
Case study, README, clear communication in multiple languages
✅ Honest Communication
Transparent about limitations and tradeoffs
The following screenshots demonstrate Atlas's capabilities using a single, end-to-end product scenario: designing a "Save for Later" feature for an e-commerce mobile app. This flow shows how Atlas supports the complete product development process from initial concept to success measurement.
Note: Atlas works identically in Spanish - simply write your prompts in Spanish and receive responses in Spanish with the same quality and structure.
Atlas greets you with clear options for the 4 main features
Input:
"Generate a user story for this e-commerce feature: Users need a way to save products..."
Output:
Well-structured user story with clear persona, action, and value proposition
Guided input collection → Calculated RICE scores → Ranked recommendations with explanations
User story input → Clear, testable acceptance criteria using Given-When-Then format
Feature description → Relevant KPIs with explanations of why each metric matters
Do not input confidential or proprietary company information into Atlas (or any public AI tool).
While Atlas is designed to help with product management tasks, it operates through Poe.com using third-party AI models (Claude). Any information you input:
- May be used to train AI models
- Is processed by external services
- Should not include sensitive business data, unreleased product details, or confidential metrics
Best practices:
- ✅ Use hypothetical scenarios and generic examples
- ✅ Anonymize any company-specific information
- ✅ Work with placeholder data for testing
- ❌ Never input actual company OKRs, roadmaps, or proprietary metrics
- ❌ Avoid real customer data or competitive intelligence
For production use with sensitive data, a custom implementation using direct API access with proper data governance would be required.
Ghiselle Butrón Reyes | Product Manager
5+ years of experience in digital product management, specializing in data-driven decision-making and cross-functional team coordination. Track record of delivering measurable results for companies like CaixaBank, Škoda, and SEAT/CUPRA.
Core Competencies:
- 📊 Data-driven product strategy
- 🤖 AI integration in product
- 🤝 Stakeholder management
- 🔄 Agile/Scrum methodologies
Professional Highlights:
- Increased lead generation by 12-30% through product optimization
- Reduced reporting time by 40% via dashboard automation
- Recent: GenAI for Product Owners certification (IBM)
Let's Connect:
- 📧 Email: ghiselle.b@gmail.com
- 💼 LinkedIn: ghiselle-butron-reyes
- 💻 GitHub: GhiselleBR
Atlas isn't revolutionary technology—it's a design exercise in making AI accessible and practical for everyday PM work.
The real innovation isn't in the code, but in recognizing that most POs don't need another powerful tool they don't know how to use. They need focused solutions that reduce friction and let them do their jobs better—in their own language.
This project demonstrates product thinking: identifying real problems, designing simple solutions, shipping quickly, and choosing the right tools for the job.
Built with ❤️ in Barcelona | November 2025