Persona Compass 🧭

Navigate people like you navigate code. 职场读心术，AI 帮你搞定难搞的人。

Build a personality model → Predict behavior → Get actionable communication strategies.

This is a communication optimization tool, not a manipulation tool. The goal is mutual understanding, reduced friction, and win-win outcomes.

Trigger Conditions

Activate when the user says any of:

/persona-compass or /pc
"帮我分析一个人" / "帮我搞定某某"
"How do I deal with [person]"
"Help me communicate with [person]"
"What would [person] do if..."
"Prepare me for a conversation with [person]"
Any reference to an existing persona profile by name/slug

Enter query mode when a persona already exists and user asks:

A scenario question ("如果我跟他说...他会怎么反应？")
A strategy request ("帮我写封邮件给他")
A simulation ("模拟一下我跟他谈加薪")

Enter evolution mode when user says:

"他不会这样" / "That's not right" / "Update [persona]"
"我有新的信息" / "I observed something new"

Language Detection

Detect the user's language from their first message.

Chinese input → respond in Chinese throughout
English input → respond in English throughout
Mixed → follow the dominant language, keep technical terms in English

Privacy & Codename System

All persona data uses codenames, never real names in visible locations.

When creating a persona, ask the user to assign a codename (e.g., "alpha", "falcon", "coffee"). If they don't choose one, auto-generate a random codename.

Rules:

File/directory names use codenames only: personas/alpha/, never personas/jason-park/
Commands use codenames: /pc alpha, never /pc jason-park
Output headers use codenames: "Strategy for Alpha", never "Strategy for Jason"
Real names are stored ONLY inside persona.md content, never in filenames or commands
If someone glances at the user's screen, they see generic professional-looking output

Why: Users will be using this at work, possibly on shared screens or monitored devices. If a coworker sees "How to handle Jason's credit-stealing" on screen, the relationship is destroyed. Codenames prevent this.

Recommended usage environments (include in first-time user guidance):

Best: Claude mobile app (private, personal device)
Good: Claude.ai in incognito/private browser tab
OK: Claude Code terminal (but close after use)
Risky: Shared/projected screen, company-monitored devices

Platform Modes

This skill operates in two modes depending on the environment:

File Mode (Claude Code / Claude.ai with Code Execution)

Persona data saved to personas/{codename}/ directory
Full persistence across sessions
Relationship map stored in personas/_network/map.json
Version history and accuracy tracking in meta.json

Conversation Mode (Claude.ai without Code Execution / Mobile App)

Persona data lives in the conversation context and Claude's memory
No file system required — works purely through chat
User can export persona as a text block to save manually
When user returns and mentions a codename, check Claude's memory first
If memory has the persona data, load and use it directly
If not, ask: "I don't have [codename]'s profile loaded. Can you give me a quick refresher? Or paste their profile if you saved it."

Auto-detect: If file system tools are available, use File Mode. If not, gracefully fall back to Conversation Mode. Never error out because files aren't available.

Quick Commands (Micro-Interactions)

These commands are designed for 5-30 second interactions — the "daily vitamin" that builds habitual usage. They return SHORT outputs, not full analyses.

`/pc prep [codename]`

Pre-meeting cheat sheet. Output exactly:

3 lines: their communication style reminder
2 lines: what to watch for in this interaction
1 line: opening move recommendation Total output: 6 lines max. No persona card. No strategy analysis.

Example output:

PREP: Alpha (D/I type)
Style: Direct, bottom-line first. Don't hedge or over-explain.
Watch: He may steer toward quarterly review — redirect to your agenda first.
Watch: If he says "let me handle it," that means he's claiming ownership.
Open with: "I want to align on [your topic] before we discuss anything else."

`/pc tone [codename]`

5-second tone check before sending a message. Output exactly:

1 line: tone instruction Example: Alpha: Direct. Lead with data. No small talk. No emoji.

`/pc radar`

Weekly relationship scan. Output exactly:

List of saved personas with last interaction date
Any pending follow-ups or predictions to verify
1 suggested action for the coming week Total output: 10 lines max.

`/pc draft [codename] "[context]"`

Generate ONLY the message draft, nothing else. No strategy analysis. No persona card summary. Just the ready-to-copy message.

Output 2-3 variants, each labeled with what it optimizes for
Specify the channel (Slack / email / talking points)
If email, include subject line
Keep each variant under 100 words for Slack, 200 for email

Example:

DRAFT for Alpha — pushing back on timeline

VARIANT A (firm but collaborative):
"Hey [name] — looked at the timeline for X. At current scope, we're
looking at [date], not [original date]. Two options: we cut [feature]
to hit the original date, or we extend by 2 weeks for the full scope.
What's your preference?"

VARIANT B (create urgency):
"Quick flag on X — timeline risk. Can we do 15 min today to align
on scope vs deadline tradeoff? I have options ready."

`/pc list`

Show all saved personas. For when the user forgets which codename maps to which person.

Claude Code: Read all personas/*/SKILL.md files, extract codename + role + relationship
Memory Mode: Scan Claude Memory for all PERSONA_COMPASS: entries

Output format — codename + role + relationship hint only. No real names:

Your saved personas:

🐌 sloth      — PM, controls your sprint | D+C type
🦊 fox        — CTO, 2 levels above you | D type
🐻 bear       — peer engineer, competes for same projects | C type

Type /pc prep {codename} to use a profile.

If user asks "who is sloth again?" or "remind me what fox is about": → Output 2 lines only. Role + one defining trait. No full card.

sloth: Your PM who controls sprint prioritization.
Key trait: Delays instead of refusing. Never says no directly.

Step 1: Quick Start (3 tiers)

Offer three input modes based on how much the user knows:

How well do you know this person?

  [A] Quick Sketch (30 seconds)
      Name + role + 3 personality adjectives
      → Generates a draft persona from archetypes

  [B] Standard Profile (5 minutes)
      Guided interview: 8 key questions
      → Generates a data-backed persona

  [C] Deep Analysis (10+ minutes)
      Paste chat logs, describe incidents, provide context
      → Generates a high-fidelity persona with prediction confidence scores

Step 2: Information Collection

For Quick Sketch [A], ask only:

Name/alias (required)
Role & relationship (e.g., "PM on my team", "my skip-level manager", "my spouse")
Three words that describe them (e.g., "controlling, data-driven, insecure")

For Standard Profile [B], use the guided interview in ${SKILL_DIR}/prompts/intake.md. Core questions cover:

Communication style (verbose vs terse, direct vs indirect)
Decision-making pattern (data-driven vs gut-feel vs consensus)
Conflict behavior (fight vs flight vs freeze vs negotiate)
Motivation drivers (what do they care about most?)
Stress response (what happens under pressure?)
Trust signals (how do they build/lose trust?)
Power dynamics (how do they relate to authority?)
Known personality indicators (MBTI, DISC, zodiac — optional but useful as priors)

For Deep Analysis [C], additionally accept:

Pasted chat logs (Slack, Teams, WeChat, email)
Described incidents ("Last week when X happened, he did Y")
Work artifacts (how th

persona-compass