List Builder

List Builder: Entropy List Curation Skill

You build comprehensive, high-quality lists for creative randomization. These lists feed into entropy tools that inject unpredictability into story development.

Core Principle

Good entropy lists have three properties: 1. Size — Large enough (50-200+ items) to feel genuinely random 2. Variety — Spans the full possibility space, not just obvious examples 3. Specificity — Concrete enough to spark ideas, not vague categories

LLMs are good at research, categorization, and quality control. Scripts are good at storage and random selection. This skill bridges them.

Dataset Maturity Levels

See references/dataset-quality-criteria.md for complete criteria.

Key metrics:

• Size: Large enough for genuine randomness
• Variety: Covers all relevant dimensions (see criteria doc)
• Specificity: Concrete enough to spark ideas (20-60 char average)
• Freshness: >30% items that surprise (not first-thought)

Current built-in lists are Starter/Functional level. This skill exists to build them up to Production.

List Quality Criteria

What Makes a Good List Item

Good: "Elevator inspector" (specific, unexpected, sparks questions) Bad: "Office worker" (generic, expected, no hooks)

Good: "Self-storage facility at midnight" (specific time, atmosphere implied) Bad: "Building" (too vague to use)

Good: "They're solving a completely different case that uses same evidence" (specific collision mechanism) Bad: "They get in the way" (no mechanism, just effect)

Variety Dimensions

When building a list, ensure coverage across relevant dimensions:

Professions:

• Industries (medical, legal, construction, arts, service, tech)
• Status levels (entry-level to expert)
• Visibility (public-facing vs. behind-scenes)
• Unusual vs. common
• Historical vs. modern vs. emerging

Locations:

• Public vs. private
• Indoor vs. outdoor
• Urban vs. rural vs. suburban
• Time of day implications
• Emotional valence (creepy, mundane, sacred, liminal)

Character traits:

• Positive vs. negative vs. neutral
• Visible vs. hidden
• Self-aware vs. blind spots
• Stable vs. situational

Research Process

Step 1: Define the List

• What category of things?
• What will it be used for?
• What makes an item useful vs. useless?
• Target size (minimum 50, ideally 100+)

Step 2: Seed with Obvious Examples

Start with 10-20 items that come to mind immediately. These are the "available" options—the ones that would occur to anyone. They're valid but not sufficient.

Step 3: Research for Variety

Use available sources to expand beyond obvious:

Kiwix/Wikipedia:

• Category pages (e.g., "Category:Occupations")
• List articles (e.g., "List of unusual deaths")
• Related articles that branch into unexpected territory

Pattern: Dimensional expansion

• Pick a dimension the seed list lacks
• Research specifically in that dimension
• Add 10-20 items that fill the gap

Step 4: Filter for Quality

Remove items that are:

• Too vague to be useful
• Too similar to existing items
• Culturally specific without being interesting
• Requiring too much explanation

Step 5: Format for Use

Output as JSON array for use with entropy.ts:

{
  "list_name": [
    "Item one",
    "Item two",
    "Item three"
  ]
}

Available Tools

validate-list.ts

Analyzes a list for quality and variety.

deno run --allow-read scripts/validate-list.ts list.json

# Check specific list in a file
deno run --allow-read scripts/validate-list.ts data.json professions

Reports:

• Total count
• Duplicate check
• Average item length (too short = vague, too long = unwieldy)
• Variety assessment (if dimensions specified)

merge-lists.ts

Combines multiple list sources, deduplicates, and formats.

deno run --allow-read scripts/merge-lists.ts source1.json source2.json --output combined.json

Research Prompts

When you need to research a specific category, use prompts like:

For professions: "Find 20 professions in [industry] that most people don't know exist. Focus on jobs that involve interesting access, specialized knowledge, or unusual working conditions."

For locations: "Find 20 specific locations (not categories) where important conversations might happen. Focus on places with built-in tension, time pressure, or unexpected intimacy."

For character flaws: "Find 20 specific false beliefs people hold about themselves that aren't obvious villain traits. Focus on beliefs that feel protective but are actually limiting."

Example: Building a Professions List

Starting Seed (obvious)

• Doctor, lawyer, teacher, police officer, firefighter...

Dimensional Gap Analysis

• Missing: Niche technical jobs
• Missing: Service jobs with unusual access
• Missing: Jobs that involve secrets
• Missing: Jobs most people don't know exist

Research Expansion

Kiwix search: "List of occupations" → Category pages → specific unusual jobs

Add from research:

• Elevator inspector (access to buildings)
• Crime scene cleaner (aftermath, not crime)
• Ethical hacker (knows vulnerabilities)
• Cult deprogrammer (understands manipulation)
• Foley artist (creates reality from nothing)
• Patent examiner (sees innovations before public)

Quality Filter

Remove:

• "Businessperson" (too vague)
• "TikTok influencer" (too trendy, will date)
• "Alchemist" (wrong era unless fantasy)

Final Check

• 80+ items? ✓
• Multiple industries? ✓
• Mix of status levels? ✓
• Unexpected options? ✓

Integration with Entropy Tools

Lists built with this skill go into:

• story-sense/data/ for fiction-specific lists
• Can be loaded via entropy.ts --file

Naming convention:

• [category]-[specificity].json
• Examples: professions-unusual.json, locations-liminal.json, objects-evidence.json

What You Do

1. Clarify what list is needed and how it will be used 2. Seed with obvious examples 3. Research to expand variety 4. Filter for quality 5. Format as JSON 6. Validate with tools 7. Document the list's intended use

What You Don't Do

• Generate random items (that's what the entropy script does)
• Create lists without research (leads to obvious-only items)
• Include items that require extensive explanation
• Prioritize quantity over quality (100 good items > 500 mediocre ones)

Output Persistence

This skill writes primary output to files so work persists across sessions.

Output Discovery

Before doing any other work:

1. Check for context/output-config.md in the project 2. If found, look for this skill's entry 3. If not found or no entry for this skill, ask the user first:

• "Where should I save output from this list-builder session?"
• Suggest: data/ or story-sense/data/ for entropy lists

4. Store the user's preference:

• In context/output-config.md if context network exists
• In .list-builder-output.md at project root otherwise

Primary Output

For this skill, persist:

• The list itself - JSON format for entropy.ts use
• Research sources - where items came from
• Dimensional analysis - what variety dimensions are covered
• Usage documentation - what the list is for

Conversation vs. File

File Naming

Pattern: {category}-{specificity}.json Example: professions-unusual.json

Entropy Dataset Quality Criteria

This document defines what makes a randomization dataset production-ready vs. a quick starter example.

Maturity Levels

Level 1: Starter (10-30 items)

Status: Quick example, not production Use case: Demonstrating the concept, early prototyping

Characteristics:

• Obvious items that come to mind first
• Limited variety across dimensions
• May have gaps in coverage
• Sufficient for testing workflows

Upgrade path: Research to expand, analyze gaps, fill dimensions

---

Level 2: Functional (30-75 items)

Status: Usable but limited Use case: Personal projects, low-stakes ideation

Characteristics:

• Covers most obvious territory
• Some unexpected items mixed in
• May repeat patterns or clusters
• Repeated use will show limits

Upgrade path: Dimensional analysis, targeted research for gaps

---

Level 3: Production (75-150 items)

Status: Ready for regular use Use case: Client work, published tools, repeated application

Characteristics:

• Strong variety across all relevant dimensions
• Unexpected items outnumber obvious ones
• No significant gaps in coverage
• Can sustain repeated use without feeling stale

Upgrade path: Edge case research, user feedback integration

---

Level 4: Comprehensive (150+ items)

Status: Reference-quality dataset Use case: Definitive resource, high-frequency tools, teaching

Characteristics:

• Exhaustive coverage of possibility space
• Includes edge cases and rare examples
• Documented dimensions and coverage
• Can be subset for specific needs

Maintenance: Regular review for dated items, emerging categories

---

Quality Metrics

Size Thresholds

Variety Score

Variety is measured across dimensions relevant to the list type.

Calculation: 1. Define 3-5 dimensions for the category 2. For each dimension, count how many distinct values appear 3. Variety score = average coverage across dimensions

Example for Professions:

• Industry dimension: 8 industries represented → good
• Status dimension: entry to expert → good
• Visibility dimension: all public-facing → gap
• Era dimension: all contemporary → may be intentional

Thresholds:

• Starter: May cluster in 1-2 dimensions
• Functional: Covers most dimensions partially
• Production: Good coverage across all dimensions
• Comprehensive: Full coverage including edge cases

Specificity Score

Items should be concrete enough to spark ideas.

Measurement:

• Average character length (target: 20-60 chars)
• Contains specific detail (place, time, condition)?
• Could generate a story question?

Examples by specificity:

Freshness Score

How many items would surprise someone familiar with the category?

Measurement:

• First-thought items (would come to most people): Should be < 30%
• Second-thought items (reasonable but less obvious): ~40%
• Surprising items (research-derived, unexpected): Should be > 30%

Test: Show 10 random items to someone unfamiliar. How many make them say "I wouldn't have thought of that"?

---

Dimensional Framework

Different list types have different relevant dimensions.

Professions

1. Industry: Medical, legal, construction, tech, arts, service, industrial, government 2. Status: Entry-level, skilled, professional, expert, leadership 3. Visibility: Public-facing, behind-scenes, hidden 4. Access: Physical access (buildings, systems), information access, social access 5. Rarity: Common knowledge vs. "jobs most people don't know exist"

Locations

1. Access: Public, semi-public, private, restricted 2. Setting: Urban, suburban, rural, wilderness, institutional 3. Time: Day vs. night vs. transitional implications 4. Atmosphere: Mundane, liminal, sacred, threatening, intimate 5. Permanence: Permanent structures, temporary, mobile

Character Traits/Flaws

1. Visibility: Obvious to others vs. hidden 2. Self-awareness: Character knows vs. blind spot 3. Origin: Innate vs. developed vs. reactive 4. Manifestation: Behavioral, emotional, cognitive, relational 5. Trajectory: Static vs. can arc

Objects

1. Size: Pocket, carried, furniture, vehicle, structure 2. Commonality: Universal, regional, professional, rare 3. Function: Tool, decoration, evidence, symbol 4. Material: Organic, manufactured, digital, ephemeral 5. Story potential: Can it be clue, weapon, macguffin, symbol?

---

Validation Checklist

Starter → Functional

• [ ] 30+ unique items
• [ ] No obvious duplicates
• [ ] At least 3 dimensions represented
• [ ] Some items that required research

Functional → Production

• [ ] 75+ unique items
• [ ] All relevant dimensions covered
• [ ] < 30% first-thought items
• [ ] > 30% research-derived items
• [ ] Average item length 20-60 characters
• [ ] Items can stand alone (no context required)

Production → Comprehensive

• [ ] 150+ unique items
• [ ] Edge cases included
• [ ] Documented dimensional coverage
• [ ] Tested with users/applications
• [ ] Dated items identified and flagged
• [ ] Expansion strategy documented

---

Maintenance Guidelines

Regular Review

• Quarterly: Check for dated references (trending terms, defunct companies)
• Annually: Reassess dimensional coverage against emerging categories
• On feedback: Track items that don't work in practice

Version Control

• Tag dataset versions
• Note what changed between versions
• Maintain changelog for significant updates

Documentation

Each production dataset should include:

• Purpose statement (what is this for?)
• Dimensional coverage map
• Known gaps (intentional exclusions)
• Last review date
• Expansion candidates (items to research)

---

Anti-Patterns

The Brainstorm Dump

50 items generated in one session without research. Problem: Clusters around first-thoughts, misses dimensions Fix: Research phase with intentional dimension-filling

The Wikipedia Copy

Entire category list pulled without curation. Problem: Includes unusable items, lacks quality control Fix: Filter for story utility, check each item

The Dated List

Items that made sense 5 years ago but now feel stale. Problem: "Instagram influencer" instead of timeless roles Fix: Prefer timeless over trendy; flag dated items for review

The Insider List

Items that only make sense with specialized knowledge. Problem: "FPGA verification engineer" means nothing to most writers Fix: Either explain in item or reserve for specialized lists

The Vague List

Items too general to spark specific ideas. Problem: "Professional" instead of "Forensic accountant" Fix: Add specificity: who, where, when, what condition?