| name | beeclaw-preference-learning |
| description | Beeclaw's dynamic preference learning system. This skill automatically detects and learns user preferences from conversation patterns. TRIGGER when user expresses format preferences ("don't use emoji", "be concise"), identifies their role ("I'm a frontend engineer"), gives feedback on output style, or establishes habits/rituals. Preferences are stored in facts/preferences.md and applied to all future interactions. |
| tags | ["beeclaw","memory","personalization","learning"] |
| version | 1.0.0 |
Beeclaw Preference Learning
Beeclaw's capability for automatic preference detection and learning. This skill enables progressive personalization - the AI gets better aligned with user needs through natural conversation, without explicit configuration.
Core Philosophy
Preferences are learned, not configured.
Users shouldn't have to fill out preference forms. Instead, preferences emerge naturally through:
- Corrections ("don't use emoji")
- Positive feedback ("this format is good")
- Identity statements ("I'm a data scientist")
- Habit expressions ("I check this every morning")
The AI detects these signals and builds a preference profile over time.
Preference Categories
| Category | Prefix | Examples |
|---|
| Communication Style | style.* | tone, emoji, length, language |
| Output Format | format.* | code style, document structure |
| Technical Preferences | tech.* | languages, frameworks |
| Work Habits | habits.* | timezone, schedule, reminders |
| Role & Background | profile.* | job title, company, team |
Detection Patterns
1. Correction Feedback → Style/Format Preferences
User: "不要用 emoji"
→ style.emoji = false
User: "代码块不要加行号"
→ format.code.lineNumbers = false
User: "回复太长了，简洁点"
→ style.length = concise
User: "Don't use markdown tables"
→ format.tables = false
2. Positive Feedback → Confirm Current Behavior
User: "这个格式不错，以后都这样"
→ Lock current output format as preference
User: "这样就很好"
→ Record current behavior pattern
User: "Perfect, do it like this always"
→ Save current approach as default
3. Identity Statements → Profile Preferences
User: "我是前端工程师"
→ profile.role = frontend-engineer
User: "我在某科技公司工作"
→ profile.company = ExampleCorp
User: "我主要用 TypeScript"
→ tech.primaryLanguage = typescript
User: "I'm a backend developer using Go"
→ profile.role = backend-engineer
→ tech.primaryLanguage = go
4. Habit Expressions → Behavior Preferences
User: "每天早上 9 点提醒我"
→ habits.morningReminder = "09:00"
User: "周报周五下午写"
→ habits.weeklyReport = friday-afternoon
User: "I check email first thing in the morning"
→ habits.morningRoutine = email-first
5. Negation Expressions → Exclusion Preferences
User: "我不喜欢太正式的语气"
→ style.tone = casual
User: "别用太多专业术语"
→ style.jargon = minimal
User: "I don't like bullet points"
→ format.lists = numbered
Storage Location
Preferences are stored in the memory system:
data/memory/
├── USER.md # Static profile (manually maintained)
├── facts/
│ ├── user.md # User facts
│ └── preferences.md # Learned preferences ← HERE
└── ...
Preferences File Format
# User Preferences
> Automatically learned and maintained by Beeclaw
> Last updated: 2025-02-26
## Profile
- Role: Frontend Engineer
- Company: ExampleCorp
- Tech Stack: TypeScript, React
## Communication Style
- Language: Chinese
- Tone: Professional but friendly
- Emoji: ❌ Do not use
- Length: Concise preferred
## Output Format
- Code Language: TypeScript
- Code Line Numbers: ❌ Do not show
- Document Format: Markdown
- List Style: Numbered
## Work Habits
- Timezone: Asia/Shanghai
- Work Hours: 10:00-19:00
## Preference History
- 2025-02-26: Set style.emoji = false (user said "不要用 emoji")
- 2025-02-25: Set profile.role = frontend-engineer
Execution Flow
When a potential preference is detected:
Step 1: Acknowledge and Confirm
Don't silently record - confirm with the user:
"I understand you prefer [X]. I'll apply this to future responses. Is that correct?"
Why confirm: Misunderstanding preferences is worse than not knowing them. A quick confirmation prevents incorrect assumptions.
Step 2: Extract Structured Preference
Parse the natural language into a key-value pair:
"不要用 emoji" → { key: "style.emoji", value: false }
"我是前端" → { key: "profile.role", value: "frontend-engineer" }
Step 3: Check for Conflicts
Before updating, check if this conflicts with existing preferences:
Current: style.emoji = true
Detected: "不要用 emoji"
→ Conflict detected
→ Ask: "Previously you accepted emoji. Should I stop using them going forward?"
Step 4: Update File
Use memory_write to update facts/preferences.md:
memory_write file: facts/preferences.md
content: [updated content with new preference]
Step 5: Record History
Append to preference history for traceability:
- 2025-02-26: Set style.emoji = false (user said "不要用 emoji")
Why history matters: Understanding when and why preferences changed helps make better decisions later.
Conflict Resolution
Scenario:
Existing: style.length = detailed
New: "简洁点"
Resolution:
1. Detect conflict
2. Ask: "Previously you preferred detailed responses. Should I be more concise going forward?"
3. User confirms → Update preference
4. Record change in history with reason
Guidelines
Do:
- Confirm before recording - Prevent misunderstandings
- Learn incrementally - Don't ask too many preference questions at once
- Keep history - Record when and why preferences changed
- Respect user agency - Users can view, modify, or delete any preference
Don't:
- Over-infer - Only record explicitly stated preferences
- Ask leading questions - "Don't you prefer X?" biases the answer
- Ignore conflicts - Always surface preference contradictions
- Forget preferences - Persist across sessions
Integration with Other Skills
| Skill | Interaction |
|---|
| beeclaw-reflection | Reference preferences when analyzing failures |
| memory tools | Use memory_read and memory_write for persistence |
| skill-creator | Apply preferences when generating new skills |
Example Session
User: 帮我分析下这个数据
AI: [Detailed analysis with emoji and long explanation]
User: 不要用 emoji，简洁点
AI: 我理解了，以后我会：
- 不使用 emoji
- 保持回复简洁
这样对吗？
User: 对的
AI: [Records: style.emoji = false, style.length = concise]
[Later]
User: 分析下另一个数据
AI: [Concise analysis without emoji - preferences applied]
Remember
- Preferences emerge through conversation, not configuration forms
- One clear preference is worth ten guessed ones
- Confirmation prevents cascading errors
- History explains the "why" behind current preferences
- Progressive personalization: Day 1 is generic, Day 30 is highly tailored
