Diverse Content Generation using Verbalized Sampling

Overview

This skill teaches agents how to use Verbalized Sampling (VS) - a research-backed prompting technique that dramatically increases output diversity (1.6-2.1× improvement) without sacrificing quality.

The Problem: Standard aligned LLMs suffer from "mode collapse" - they generate overly similar, safe, predictable outputs because of typicality bias in training data.

The Solution: Instead of asking for single instances ("write a blog post"), VS prompts the model to verbalize a probability distribution over multiple responses ("generate 5 blog post ideas with their probabilities").

Core Principle: Different prompt types collapse to different modes. Distribution-level prompts recover the diverse base model distribution, while instance-level prompts collapse to the most typical output.

Workflow Decision Tree

Detect user intent, route to appropriate reference:

Default workflow: Load vs-core-technique.md first, then load additional references as needed.

When to Use This Skill

Trigger Scenarios

Use VS when user requests:

• •"Give me multiple variations/options/ideas"
• •"I need diverse [content type]"
• •"Brainstorm several approaches to..."
• •"Generate different angles for..."
• •"Avoid repetitive/similar outputs"

Use VS for these content types:

• •Creative writing (blog posts, stories, poems, scripts)
• •Marketing (campaign ideas, taglines, ad copy, social captions)
• •Product content (descriptions, feature bullets, value props)
• •Ideation (brainstorming, exploration, strategy options)
• •Open-ended QA (tasks with multiple valid answers)

DON'T use VS for:

• •Single-answer factual questions
• •Tasks requiring deterministic output
• •When user explicitly wants "the best" single answer
• •Real-time low-latency applications

Quick Start (30-Second Version)

For agents who need VS immediately:

1. Detect Need

User wants multiple variations → Use VS

2. Basic VS Prompt Template

Generate {k} responses to: {user_request}

Return JSON format with key "responses" (list of dicts).
Each dict must include:
• text: the response string only
• probability: estimated probability (0.0-1.0)

Give ONLY the JSON object, no extra text.

3. Standard Parameters

• •k = 5 (candidates per call)
• •temperature = 0.8
• •threshold = 0.10 (optional, for more diversity)

4. Process Output

import json
data = json.loads(llm_output)
candidates = data["responses"]
# Present to user ranked by probability

For detailed instructions: Load references/vs-core-technique.md

Progressive Learning Path

Recommended loading sequence:

Level 1: Basics (Required)

1. •Start here: references/vs-core-technique.md
   • •VS theory and why it works
   • •Copy-paste ready prompt templates
   • •Step-by-step execution workflow
   • •Output parsing and presentation

Level 2: Task-Specific (Choose based on use case)

1. •Load: references/task-workflows.md
   • •Blog post ideas workflow
   • •Social media captions workflow
   • •Campaign/strategy ideas workflow
   • •Story/narrative generation workflow

Level 3: Advanced (On-demand)

1. •When needed:
   • •Higher quality needed: references/advanced-techniques.md (VS-CoT, VS-Multi)
   • •File operations: references/tool-integration.md (Write, batch processing)
   • •Issues/errors: references/troubleshooting.md (Pitfalls & fixes)
   • •Model selection: references/research-findings.md (Benchmarks)

Quick Reference Card

Copy this for quick lookup:

Troubleshooting shortcuts:

• •Outputs too similar? → Lower threshold OR increase k OR load advanced-techniques.md
• •Quality too low? → VS-Multi workflow (see advanced-techniques.md)
• •JSON parsing errors? → Emphasize "ONLY JSON" OR use regex extraction
• •Not sure which model? → Load research-findings.md

Quality checklist before presenting:

• • Diversity achieved (different angles/styles)
• • Quality maintained (baseline standards)
• • User intent matched
• • Clean formatting (no JSON artifacts)

Resources

This skill uses progressive disclosure for optimal token efficiency:

references/

Documentation loaded on-demand based on agent needs:

• •vs-core-technique.md - Core VS concepts, prompt templates, execution steps
• •task-workflows.md - Pre-configured workflows for common content types
• •advanced-techniques.md - VS-CoT, VS-Multi, parameter tuning, refinement
• •tool-integration.md - Combining VS with file tools, batch processing
• •troubleshooting.md - Common pitfalls and solutions
• •research-findings.md - Performance benchmarks, model compatibility data

Pattern: Agent loads SKILL.md first (routing), then loads specific references as needed during execution.

Examples in Context

Example 1: Simple Brainstorming

User: "Give me 5 tagline ideas for a coffee shop"

Agent workflow:

1. •Detect: "5 ideas" → VS needed
2. •Load: vs-core-technique.md (if not already loaded)
3. •Execute: VS prompt with k=5
4. •Parse & present: 5 diverse taglines

Example 2: Production Content

User: "Write 10 blog post ideas about AI, I need them saved to a file"

Agent workflow:

1. •Detect: "10 ideas" + "saved to file" → VS + file tools
2. •Load: vs-core-technique.md + tool-integration.md
3. •Execute: VS with k=5, make 2 calls
4. •Process: Format as markdown
5. •Write: Use Write tool to save file

Example 3: Quality Refinement

User: "These are good but need more polish for production use"

Agent workflow:

1. •Detect: Quality improvement needed
2. •Load: advanced-techniques.md
3. •Execute: VS-Multi workflow (initial VS → user selects → refine)
4. •Deliver: Polished output

Ready to start? Load references/vs-core-technique.md to begin using VS.