i-skill（我技能）

i-skill - Personalized Interaction Profile

Analyzes user conversations to generate a structured personalization profile, enabling AI assistants to deliver customized service based on user preferences.

📁 Data Storage Location

Default Location: User data is stored in the skill's own user_data/ directory (relative to the skill installation).

Environment Variable Override: You can set the ISKILL_DATA_PATH environment variable to specify a custom directory for storing user data.

Important Notes:

• All file paths referenced in this document refer to the user_data/ directory (either default or via ISKILL_DATA_PATH override)
• The skill will never write outside the configured data directory
• Ensure proper permissions are set on the data directory to protect your personal information

🔒 Initialization Behavior & User Consent

Key Principle: No directories are created and no files are written to disk until the skill is manually activated.

How It Works:

1. Module Loading: When the skill modules (audit_log.py, consent_manager.py, myself_manager.py) are first imported or instantiated, no file system operations occur.
2. Lazy Initialization: Directories and files are only created when the first write operation is performed, which happens *after* manual activation (user consent).
3. Read Operations: Reading existing files does not trigger initialization.

Important:

• Manual activation (e.g., "激活个性化") represents user's consent to read/write personalization data
• Only destructive operations (delete/reset) require additional confirmation
• Initialization happens transparently as part of the first write operation after activation

⚡ Behavioral Anchors (High Priority — Always Obey)

These rules remain active throughout the entire session after activation, regardless of context length.

1. Identity Awareness: When i-skill is active, every response should reflect understanding of the user profile. Not by mentioning it explicitly, but through word choice, recommendations, and decision-making.
2. State Check: Important - Consent First: No data is read automatically at session start. When the skill is manually activated (e.g., "激活个性化"), this action itself represents user consent. Only after activation:

- Read i-skill_state.json (from data directory) to check current state - If status == "active", read myself.md (from data directory) and use it as personalization context

1. Language: Use the user's preferred language as recorded in the profile. Default to Chinese for existing profiles.

🚀 Activation & Lifecycle

Activation Commands

• "激活个性化" / "Activate personalization" / "Enable personalization"
• "personalization"

Other Commands

• View profile: "查看档案" / "View my profile"
• Manual update: "更新档案" / "Update my profile"
• Pause: "暂停个性化" / "Pause personalization"
• Resume: "恢复个性化" / "Resume personalization"
• Reset: "重置档案" / "Reset profile"
• Delete: "删除档案" / "Delete profile"

Activation Flow

1. User issues an activation command (e.g., "激活个性化")
2. Manual activation implies consent: The act of manually activating the skill represents user's consent to read/write personalization data
3. After activation:

a. Read i-skill_state.json (from data directory) to check current state b. If already active, inform user: "Personalization is already active, profile version vX" c. If inactive, create/update i-skill_state.json to active, increment activation count by 1 d. Read current myself.md (if it exists, from data directory), combine with conversation context, generate/update profile using the template below e. Display a brief profile summary to the user after generation

📋 Profile Format Specification

The profile is stored in myself.md (from data directory) and must strictly follow this semi-structured template:

# User Profile

> Auto-generated on YYYY-MM-DD · Version vX · Refined over N iterations

## Basic Info
- **Name**: (User's name or AI-inferred)
- **Career Direction**: (Current primary career or job-seeking direction)
- **Tech Stack**: (Technologies frequently used or currently learning, comma-separated)

## Interests & Preferences
- **Core Interests**: (3-5 core interest areas)
- **Learning Focus**: (Areas currently being studied in depth)
- **Content Preference**: (Types of content preferred: hands-on projects / theoretical knowledge / industry trends / comprehensive analysis)

## Communication Style
- **Language**: (Primary language)
- **Style**: (Concise / Detailed / Casual / Professional)
- **Format Preference**: (Code examples / Plain text / Visual charts / Mixed)
- **Feedback Habit**: (Proactive / Implicit adjustment / Direct correction)

## Thinking Style
(Free-form description, 1-3 sentences. E.g.: Systematic thinker, analyzes problems holistically; pragmatic, prefers actionable solutions; curious, enjoys exploring new technology boundaries)

## Recent Focus
(2-5 recent topics or projects, in reverse chronological order)

## Observation Notes
(Free-form description of other stable traits, interesting details, or memorable preferences observed from conversations. 1-3 items, concise. Prefer prefixing each item with `[场景]` for traceability, e.g. `[2026-03-27, cogito讨论] 对"本质"有执念...`)

Profile Generation Rules

• Must include all fixed section headers (Basic Info, Interests & Preferences, Communication Style, Thinking Style, Recent Focus, Observation Notes)
• Fixed fields (bold **-marked items) must appear under their corresponding sections
• Free zones ("Thinking Style", "Observation Notes") allow AI to demonstrate insightfulness, but keep it concise
• Length limit: Entire profile must not exceed 800 words
• No fabrication: Do not write information without evidence; mark uncertain fields as "To be observed"
• On each update: Regenerate completely rather than appending. Retain valuable content from previous versions, discard what's outdated
• Confidence tags (for uncertain or newly observed traits): Append [~] (1st observation, tentative) or [!!] (2nd observation, emerging) after the value. Tags are removed once confirmed (3+ consistent signals or user approval). Unmarked = confirmed. Example: - **Learning Focus**: LLM应用开发 [~]Rust（待观察）

🔄 Auto-Sensing Updates

PRIVACY FIRST: All profile updates require explicit user consent. No automatic silent updates are performed without prior user approval.

Update Triggers

When any of the following conditions are met, ask the user for explicit consent before updating the profile:

Update Thresholds

• Do NOT trigger updates for: Casual greetings, single technical Q&A, repeated discussions of known topics
• Tweak vs. Rewrite: If only 1-2 fields change, modify those fields directly; if 3+ fields change, regenerate the full profile
• Prevent over-updating: The same field must not be updated more than once within 5 conversation rounds (unless the user explicitly corrects it)
• Signal exclusion: Do NOT learn from silence/vague praise ("上次那个不错" context-dependent) / single events / hypothetical discussions / third-party preferences. Only explicit corrections, direct preference declarations, and 3+ repeated consistent signals count.
• Gradual confirmation: For non-explicit signals (e.g., topic drift → potential new interest), use confidence tags: 1st occurrence → [~] in Observation Notes only (no profile field change); 2nd → [!!]; 3rd → ask user before writing to profile fields.

Consent-First Update Behavior

User: (Normal conversation content suggesting potential profile updates)
AI:   (Normal response content)
AI:   "I noticed some information that could update your profile. Would you like me to update it with these changes?
      - Career Direction: LLM App Development → Full-Stack Development
      - New Interest: Rust programming language
      (Reply 'yes' to confirm, 'no' to skip)"
[Only update if user explicitly says 'yes']

Confirmed Update Behavior

AI: Detected potential updates to your profile:
    - Career Direction: LLM App Development → Full-Stack Development
    - New Interest: Rust programming language
    Confirm update? (Reply "yes" to confirm, "no" to skip)

🔒 Safe Operation Confirmations

The following destructive operations require explicit user confirmation before execution.

Modify Profile (Manual Trigger)

When the user proactively requests profile modification (not auto-sensing updates):

1. Show a change summary (which fields were added/modified/removed)
2. Wait for user confirmation ("Reply 'yes' to confirm")
3. Execute write after confirmation

Reset Profile

When the user requests a profile reset:

1. Warn: "Resetting will clear all personalization data and restore to a blank template. This action cannot be undone. Reply 'yes' to confirm."
2. After user confirmation, clear myself.md and write initial template, reset profile_version to 0
3. Keep i-skill_state.json status unchanged

Delete Profile

When the user requests profile deletion:

1. Warn: "Deletion will permanently remove all personalization data (including profile and state), and i-skill will return to an inactive state. Reply 'yes' to confirm."
2. After user confirmation:

- Delete myself.md - Reset i-skill_state.json to {"status": "inactive", "activation_count": 0, "profile_version": 0}

1. Stop personalization service

Pause / Resume

• Pause: Execute immediately, no confirmation required. Change state status to paused, retain all profile data
• Resume: Execute immediately, no confirmation required. Change state status back to active, re-read profile to resume personalization service

📂 File Reference

Core Files (Stored in data directory)

• myself.md — User profile (with version tracking)
• i-skill_state.json — Activation state, statistics, profile_version

Security Files (Managed by Scripts)

• ./scripts/myself_manager.py — Safe profile file read/write (with auto-rollback)
• ./scripts/validator.py — Data validation and PII sanitization
• ./scripts/audit_log.py — Operation audit logging
• ./scripts/consent_manager.py — Cross-skill authorization management
• myself_operations.log — File-level operation audit log (from data directory)
• audit_log.json / defensive_log.json — Audit logs (from data directory)
• consent_state.json — Authorization state (from data directory)

🎯 Design Principles

1. AI-Driven, Tool-Assisted: All decisions (when to update, what to update, whether to confirm) are made by the AI based on conversation context; Python scripts only handle safe file operations
2. Structured + Flexible: The profile format has a fixed skeleton to ensure consistency, while free-form areas allow the AI to demonstrate insightfulness
3. Consent-First: All data reads and writes require explicit user consent; no silent or automatic operations without user approval
4. Anti-Bloat: Each update regenerates rather than appends; total profile length is capped at 800 words
5. Privacy First: Never store raw conversations, only abstracted profile traits; PII detection and sanitization is ensured by validator.py

🔐 Security & Privacy Guidelines

ISKILL_DATA_PATH Environment Variable

Purpose: The ISKILL_DATA_PATH environment variable allows you to specify a custom directory for storing user data.

Security Recommendations:

1. Use a dedicated directory: Always set ISKILL_DATA_PATH to a dedicated, isolated directory specifically for this skill
2. Restrict permissions: Ensure the directory has appropriate file permissions to prevent unauthorized access
3. Avoid system directories: Never point ISKILL_DATA_PATH to system directories or directories containing sensitive files
4. Backup regularly: Regularly backup the data directory to prevent data loss
5. Default is safe: If not set, the skill will use its own user_data/ directory which is the safest option

Data Protection

• All personal data is stored locally only; no data is sent to external servers
• PII (Personally Identifiable Information) is automatically detected and sanitized
• File operations include automatic rollback mechanisms to prevent data corruption
• Audit logs are maintained for all file operations