Agents Prompt Library

​

The Problem

• No version history you can actually use — Google Docs tracks character-level edits, not meaningful prompt versions. Rolling back to “the version that worked well last Tuesday” is a nightmare.
• No structure — Prompts have variables like {{user_name}}, {{language}}, {{loan_amount}} but there’s no way to test them with real values without copy-pasting into a playground.

​

Your Mission

1. Create and organize prompt templates with metadata
2. Fill in variables and preview the rendered prompt
3. Test prompts against an LLM and see the response instantly
4. Track versions so we know what changed and when

​

Time Budget

​

Technical Requirements

​

Database Layer (PostgreSQL)

prompts
├── id (uuid)
├── name (e.g., "Loan Reminder Agent — Hindi")
├── description
├── category (e.g., "collections", "onboarding", "survey", "verification")
├── language (e.g., "hi", "en", "es", "mr")
├── template_body (the prompt text with {{variables}})
├── variables (JSON array of variable definitions)
├── version (integer, auto-incrementing per prompt)
├── is_active (boolean)
├── created_at
└── updated_at

prompt_versions
├── id (uuid)
├── prompt_id (foreign key)
├── version (integer)
├── template_body (snapshot of the prompt at this version)
├── change_note (e.g., "Made tone more polite for rural users")
├── created_by
└── created_at

​

Backend Layer

• GET /prompts — List all prompts (with filters: category, language, search by name)
• GET /prompts/:id — Get a single prompt with its version history
• POST /prompts — Create a new prompt
• PUT /prompts/:id — Update a prompt (auto-creates a new version)
• DELETE /prompts/:id — Soft delete (set is_active = false)

• POST /prompts/:id/render — Accept variable values, return the rendered prompt with all {{variables}} filled in
• POST /prompts/:id/test — Render the prompt, send it to an LLM API, return the model’s response

• GET /prompts/:id/versions — List all versions of a prompt
• GET /prompts/:id/versions/:version — Get a specific version
• POST /prompts/:id/rollback/:version — Rollback to a previous version (creates a new version with the old body)

​

Frontend Layer (React or HTML + JS)

• Searchable, filterable table/grid of all prompts
• Show name, category, language, version number, last updated
• Filter by category and language

• Text editor for the prompt template body
• Auto-detect {{variables}} from the template and show input fields for each
• Live preview panel showing the rendered prompt as you type variable values
• “Save” button that creates a new version with an optional change note

• Fill in variables → click “Test with LLM” → see the model’s response
• Show latency and token count of the LLM response

• List of past versions with timestamps and change notes
• Click to view any version
• “Rollback to this version” button

​

Example Prompt Template

You are a polite loan reminder agent for {{company_name}}.
You are calling {{user_name}} regarding their pending EMI payment.

Details:
- Loan ID: {{loan_id}}
- EMI Amount: ₹{{emi_amount}}
- Due Date: {{due_date}}
- Days Overdue: {{days_overdue}}

Instructions:
- Speak in Hindi (Hinglish is okay)
- Be polite but firm
- If the user says they will pay, ask for a specific date
- If the user says they cannot pay, ask about their situation empathetically
- Never threaten or use aggressive language

[
  {"name": "company_name", "type": "string", "default": "Awaaz Financial"},
  {"name": "user_name", "type": "string"},
  {"name": "loan_id", "type": "string"},
  {"name": "emi_amount", "type": "number"},
  {"name": "due_date", "type": "date"},
  {"name": "days_overdue", "type": "number"}
]

​

Constraints & Ground Rules

​

Evaluation Criteria

​

Bonus Challenges 🏆

• Diff View: Show a side-by-side diff between two prompt versions (use a library like diff-match-patch or jsdiff)
• Prompt Duplication: “Clone this prompt” button to quickly create a variant (e.g., same prompt but for a different language)
• Tag System: Add freeform tags to prompts for more flexible organization
• Response Rating: After testing with an LLM, let the user rate the response (👍/👎) and store it — helps track which prompt versions perform best
• Export: Export a prompt as JSON or Markdown for sharing