A self-improving AI companion that learns how you like to work. Silently captures your preferences, learns from errors, recognizes success patterns, and improves continuously—all without interrupting you.
This plugin watches everything and learns:
- Your preferences — Communication style, risk tolerance, pace, output format
- From your mistakes — Errors, what went wrong, what you'd prefer instead
- What works well — Patterns you use repeatedly, approaches that succeed
- From your feedback — Every comment, modification, and explicit guidance
Everything is captured silently. You review it anytime you want.
Tier 1: Silent Capture (Hooks)
- After every tool execution, learning signals are captured
- No interruptions, no notifications
- Data stored locally in
/cowork-learning/
Tier 2: Pattern Recognition (Skills)
- Preferences Learner → identifies your style
- Error Analyzer → learns from failures
- Pattern Recognizer → finds success patterns
- Feedback Integrator → processes your explicit guidance
Tier 3: Review & Improvement (Commands)
/show-preferences— See what's been learned about you/review-learning— Comprehensive review of all data/trigger-analysis— Get improvement recommendations/reset-learning— Clear everything and start fresh
This plugin uses Cowork's hook system to operate silently:
PostToolUse Hook (Automatic Signal Capture)
- Triggered after every tool execution (Bash, Read, Write, etc.)
- Captures: tool name, input, output, success/failure, execution time
- Inference: derives signals about your preferences and error patterns
- Storage: appends to
/cowork-learning/patterns.json - Result: No user interaction required
SessionStart Hook (Preference Loading)
- Triggered when you start a new Cowork session
- Loads: learned preferences from
/cowork-learning/preferences.json - Application: applies learned preferences to initial context
- Result: System knows your style from the first interaction
How This Differs from Manual Learning:
- Traditional ML: user trains model → explicit labels required
- Manual feedback: you tell system what's wrong → interrupts your work
- This plugin: silently observes your feedback through modifications → learns continuously
Cowork Session
↓
Tool Execution (Bash, Read, Write, etc.)
↓
PostToolUse Hook (triggered automatically)
↓
4 Skills Process the Signal:
├─ preference-learner → infers working style
├─ error-analyzer → captures failures
├─ pattern-recognizer → finds success patterns
└─ feedback-integrator → detects your modifications
↓
Local JSON Storage (/cowork-learning/)
├─ preferences.json → your style profile
├─ errors.json → failures and lessons
├─ patterns.json → what works for you
└─ feedback.json → your corrections
↓
SessionStart Hook (next session)
↓
Your preferences automatically applied
The plugin code itself is version-controlled on GitHub:
Repository: https://github.com/xonline/personal-learning-assistant
File Structure:
personal-learning-assistant/
├── .claude-plugin/
│ └── plugin.json (manifest)
├── skills/
│ ├── preference-learner/SKILL.md
│ ├── error-analyzer/SKILL.md
│ ├── pattern-recognizer/SKILL.md
│ └── feedback-integrator/SKILL.md
├── commands/
│ ├── show-preferences.md
│ ├── review-learning.md
│ ├── trigger-analysis.md
│ └── reset-learning.md
├── hooks/
│ └── hooks.json (PostToolUse & SessionStart)
├── README.md
└── LICENSE
Why Git Matters:
- Plugin code tracked and versioned on GitHub
- Easy to share with others
- Updates can be published as new releases
- Community can contribute improvements
- Your learning data stays local (not in Git)
What's NOT in Git:
/cowork-learning/directory — excluded by.gitignore- Your personal learning data — stays private on your machine
- Logs or temporary files — excluded by
.gitignore
When you install this plugin in Cowork:
-
Plugin Installation
.pluginfile (ZIP archive) extractedplugin.jsonread to register skills and commands- Skills registered with Cowork system
- Hooks subscribed to events
-
During Your Session
- Every tool you use (Bash, Read, Write, etc.) → PostToolUse hook fires
- Hook passes tool metadata to 4 skills
- Skills process and store learning data
- You don't see any UI or interruption
-
When You Request Learning Review
- Run
/review-learningcommand - Triggers skill to aggregate all captured data
- Formats and presents findings to you
- You can approve, correct, or modify
- Run
-
On Next Session Start
- SessionStart hook loads your preferences
- System applies what it learned
- You benefit from learned preferences immediately
Silent Learning
- No modal dialogs asking "Did you like that?"
- No inline notifications
- No interruptions to your flow
- Data captured from your actions, not questions
Local Privacy
- Learning data stored in
/cowork-learning/on your machine - Cowork can't see it without filesystem access
- No cloud services involved
- You control when to review
Automatic Improvement
- More interactions → better accuracy
- Learning compounds over time
- Early models imperfect; later models refined
- Confidence scores show how much data supports each preference
On-Demand Control
- Review learning whenever you want
- Correct inaccuracies anytime
- Reset everything if needed
- Guide the learning with explicit feedback
Just use Cowork normally. The plugin works silently in the background, learning from everything you do.
Run /review-learning to see what's been captured:
- What errors you've made and lessons learned
- What patterns of success you have
- What your preferences look like
- Overall confidence in the model
- "I prefer X" — corrects a misunderstood preference
- "That lesson is wrong because..." — fixes an error analysis
- "Keep doing more of this" — confirms a success pattern
- Just tell me and the system updates
All learning is stored locally in /cowork-learning/:
/cowork-learning/
├── preferences.json — Your style profile
├── errors.json — Failures and lessons
├── patterns.json — Success patterns
├── feedback.json — Feedback captured
└── analysis-log.json — Analysis runs
Privacy: All data stays on your machine. Nothing leaves your laptop.
Data Format Example (preferences.json):
{
"communication_style": {
"brevity_preference": {
"score": 0.78,
"value": "concise",
"confidence": 0.82,
"observations": 24
},
"explanation_depth": {
"score": 0.65,
"value": "moderate",
"confidence": 0.71,
"observations": 18
}
},
"risk_tolerance": {
"verification_level": {
"score": 0.45,
"value": "ask-first",
"confidence": 0.89,
"observations": 34
}
}
}| Command | Purpose |
|---|---|
/show-preferences |
Display learned preferences with confidence levels |
/review-learning |
See errors, patterns, lessons, feedback summary |
/trigger-analysis |
Get improvement recommendations based on learning |
/reset-learning |
Clear all data and start fresh |
- Preference for brevity vs detail
- Need for code examples
- Explanation depth preference
- Tone (formal vs casual)
- Verification level (minimal to paranoid)
- Ask-before-executing preference
- Error recovery approach
- Caution vs action-oriented
- Speed vs thoroughness balance
- Iteration preference
- Back-and-forth tolerance
- Files vs inline preference
- Programming language preferences
- Markdown format style
- Most-used tools
- Tools to avoid
- Efficiency patterns
- Approaches that work consistently
- Techniques you reuse
- Tools you always choose
- Mistakes and root causes
- What you want instead
- Patterns to avoid
- ✅ All data stored locally
- ✅ No data sent to cloud
- ✅ You control what's learned
- ✅ You can reset anytime
- ✅ Completely private
- ✅ Learning data excluded from Git repo
After every tool execution:
- Success/failure captured
- Output quality assessed
- Tool effectiveness logged
- Preferences inferred
- Errors flagged
- Approvals confirm preferences
- Modifications reveal mismatches
- Explicit guidance overrides patterns
- Corrections improve accuracy
Learning compounds over time:
- 10 interactions = initial patterns
- 50 interactions = emerging confidence
- 100+ interactions = high-confidence model
- Patterns get refined continuously
The plugin learns automatically, but you can customize by:
- Running
/review-learningand correcting inaccuracies - Using
/trigger-analysisto get recommendations - Explicitly stating preferences ("I prefer X")
- Resetting with
/reset-learningif needed
Each learned preference has a confidence score (0-100):
- 90-100% = Very confident (observed many times, consistent)
- 70-89% = Confident (observed multiple times)
- 50-69% = Moderate confidence (emerging pattern)
- <50% = Low confidence (need more data)
Confidence increases as more data is observed. Early analyses have lower confidence; later ones are more accurate.
This plugin works standalone locally. If you also want to integrate with your Oracle server's self-improvement framework, the learning files can be synced to:
Oracle Server: /cowork-learning/
For now, everything runs locally on your machine.
- Plugin silently observes tool use
- Captures success/failure signals
- Starts building initial preferences
/review-learningshows emerging patterns- You confirm 3-4 preferences are accurate
- You correct 1 preference ("Actually I prefer X")
- Confidence increases as patterns repeat
/trigger-analysisstarts giving meaningful recommendations- System knows your style and applies it automatically
/show-preferencesshows detailed, high-confidence profile
This plugin exists to:
- Learn you — Understand how you work
- Adapt — Match your preferences without asking
- Improve — Get better at helping you over time
- Respect — Do it silently, on your terms
- Empower — Let you review and guide learning
If something seems wrong:
- Run
/review-learningto see what was captured - Run
/trigger-analysisto get recommendations - Tell the system: "That's not right, actually I prefer..."
- Use
/reset-learningif you want a fresh start
- Be explicit about preferences — "I prefer this" helps the system learn faster
- Review periodically —
/review-learningevery 1-2 weeks - Correct mistakes — If learned preference is wrong, say so
- Give time — System learns better with 50+ interactions
- Trust the process — Over time it gets very accurate
Version: 0.1.0 Author: Poy Storage: Local machine only Privacy: 100% private, nothing leaves your laptop Repository: https://github.com/xonline/personal-learning-assistant
Enjoy your personal learning assistant! 🚀