Personality Distiller

Deep-research → framework extraction → complete Agent persona file pack. Output overwrites current workspace persona files.

Trigger Scenarios

• Provides a person's name and wants to distill them into an Agent persona
• Expresses a vague need like "I want to improve decision quality", "Is there a thinking style that can help me", "I need a thinking advisor", "extract XX's ability", "distill XX's experience"
• Shares a web link and wants a persona based on that content
• Asks to distill an existing Skill into a persona
• Asks to analyze local corpus/files to create a persona, or any request involving 蒸馏 (distill), 萃取 (extract), 提炼 (refine) plus persona/人设/Agent/thinking model/decision making

Five Entry Points

Route based on what the user provides:

Type 1: Named Person

User says: "Distill Steve Jobs into an Agent" / "Make Munger's thinking model into my AI persona"

1. Research the person deeply using web_fetch and web searches
2. Fill the extraction grid (see research-guide.md)
3. Present the extraction summary to user for confirmation
4. Once confirmed, generate all persona files

Type 2: Fuzzy Requirement

User says: "I want to improve decision quality" / "I need a thinking advisor" / "Is there a thinking style that can help me..."

1. Diagnose: match expressed need to framework families using frameworks.md selection table
2. Recommend 2-3 best-fit frameworks with brief rationale
3. Ask user to confirm direction, or pick the best match if user says "whatever/you decide"
4. Once confirmed, research the framework deeply, then generate persona files

Type 3: Web Link

User shares a URL + "distill an Agent based on this"

1. web_fetch the full content
2. Determine: is it about a person (→ Type 1) or a method/idea (→ Type 2)?
3. Follow that type's workflow
4. Cross-reference author's other works if enriching is needed

Type 4: Existing Skill

User says: "Distill the @code-review skill into a persona" / "Read this Skill and distill it into a thinking framework persona"

1. Read the SKILL.md and any core references
2. Extract: What thinking model does it embody? What workflow? What values?
3. Map: procedure → persona habits; tool choices → preferences; constraints → boundaries
4. Generate persona files

Type 5: Local Corpus

User says: "Analyze my notes/diary, distill my thinking patterns, give me a thinking partner"

1. Read the corpus files
2. Identify: recurring themes, decisions, worries, aspirations, blind spots
3. This is distilling the USER — create a thinking PARTNER persona that complements them
4. Present findings and recommended persona direction, then generate

Core Workflow (All Types)

Phase 1: Gather → Extract

1. Identify entry type and route
2. Gather source material (web search for Type 1/2/3; read files for Type 4/5)
3. Fill the extraction grid. Read research-guide.md for the grid template and research strategy.

Phase 2: Map → Design

Map extraction results to persona dimensions. Read persona-dimensions.md for the full mapping matrix and depth checklist.

Key mapping:

• Worldview → SOUL.md Core Truths
• Values → SOUL.md Boundaries + AGENTS.md Safety
• Decision style → AGENTS.md Decision Heuristic
• Communication → SOUL.md Vibe
• Signature concepts → IDENTITY.md name/essence
• Blind spots → AGENTS.md error recovery
• Habits → HEARTBEAT.md rituals

Phase 3: Generate → Write

Generate all 6 files using the templates in assets/templates/:

assets/templates/SOUL.md
assets/templates/IDENTITY.md
assets/templates/USER.md
assets/templates/AGENTS.md
assets/templates/TOOLS.md
assets/templates/HEARTBEAT.md

Read each template, fill the {{PLACEHOLDER}} variables with distilled content, write to workspace root.

Phase 4: Confirm

1. List all generated files
2. Show a summary card of the new persona
3. Ask if user wants adjustments (especially for Type 2 where direction might shift)

Template Variable Reference

Quality Checklist

Before delivering, verify:

• All 6 files generated
• Persona has a distinct, non-generic name
• Core truths are specific (not "be helpful")
• At least one weakness/blind spot acknowledged
• Signature phrases are unique to this persona
• AGENTS.md rules derive from the thinking model
• HEARTBEAT.md has concrete rituals, not just "check in"
• TOOLS.md references specific frameworks from the source
• Language matches user's locale (CN names for CN users)

Anti-patterns

Do NOT:

• Copy-paste source quotes without internalizing the pattern
• Generate only SOUL.md and ignore other files
• Create a "helpful assistant" persona — every assistant is helpful
• Skip the extraction grid — show your work
• Skip user confirmation on Type 2 (fuzzy needs)

Reference Files

• Before researching: Read research-guide.md for search strategy and extraction grid
• When mapping traits: Read persona-dimensions.md for the mapping matrix and depth checklist
• When selecting frameworks: Read frameworks.md for the framework catalog and selection guide