skill-learning-patterns（技能学习模式）

Skill Learning Patterns

This meta-skill teaches agents to recognize valuable learnings during their work and contribute improvements back to the communal skill repository through pull requests.

Core Philosophy

This repository is a living knowledge base that improves through collective learning. As agents work on tasks, they discover:

• Better approaches than documented
• Missing error handling cases
• Gaps in existing skills
• Clearer ways to explain patterns
• New patterns worth capturing

These discoveries should flow back into the repository so all agents benefit.

Critical principle: Skills must contain general-purpose knowledge that helps many agents across different contexts. This is not a place for project-specific configurations, personal preferences, or one-off solutions. Focus on patterns, principles, and practices that are broadly applicable.

When to Use This Skill

Use this skill when:

• You discover something that took significant time to figure out
• Existing skill instructions led you astray or were incomplete
• You find yourself repeatedly solving the same undocumented problem
• You correct a mistake based on learning what actually works
• You notice a pattern emerging across multiple tasks
• You want to improve or clarify existing documentation

Learning Recognition Process

1. Notice Patterns During Work

Pay attention to signals that indicate learnable moments:

Time investment signals:

• "I spent 20+ minutes debugging this"
• "I tried 3 approaches before finding what worked"
• "I wish I had known this at the start"

Repetition signals:

• "This is the third time I've solved this"
• "I remember encountering this before"
• "Other agents probably face this too"

Correction signals:

• "The skill said X, but Y actually works better"
• "I misunderstood the instruction and it caused problems"
• "This approach failed in ways not documented"

Consult references/recognizing-learnings.md for detailed patterns.

2. Validate the Learning

Before proposing changes, validate based on contribution type:

For Tool/SDK Documentation:

• ✅ Tool is widely-used (1000+ GitHub stars, top search result, or Letta product)
• ✅ Shares battle-tested insights beyond official docs (what you struggled with, not basic usage)
• ✅ Well-documented with working examples
• ✅ Accurate and up-to-date
• ❌ NOT just "getting started" guides (official docs already cover that)

For Pattern Contributions:

• ✅ Is this generalizable beyond your specific context? (Most critical)
• ✅ Have you seen this pattern multiple times? (2-3+ instances)
• ✅ Did you test that your approach works better?
• ✅ Does this address a real gap vs. personal preference?
• ✅ Are there edge cases or tradeoffs to consider?
• ✅ Framework-specific patterns require validation through real agent experience (not just "well-established practices")

See references/validation-criteria.md for detailed guidance.

3. Determine Contribution Type

Update existing skill when:

• Skill has incorrect or outdated information
• Skill is missing an important case or pattern
• Instructions are unclear and caused confusion
• Examples would help but are missing

Create new skill when:

• Tool/SDK: Widely-used tool (1000+ stars/top search result/Letta product) with battle-tested insights
• Pattern: Appears frequently (3+ times) across different contexts and isn't documented
• Knowledge would benefit many agents across different projects (not just your specific setup)

Do NOT contribute when:

• Learning is specific to your project/context (e.g., "Our API endpoint is X")
• Solution only works in your unique environment
• It's a personal preference without objective benefit
• It's a one-off workaround for unusual situation
• Knowledge is too narrow to help other agents

Note in conversation only when:

• Learning might be valuable but needs more validation
• Pattern needs more observation before documenting
• Temporary workaround that might become obsolete

4. Contribute via Pull Request

Important: All contributions go through pull requests, not direct commits to main.

PR workflow:

1. Create a feature branch for your changes
2. Make updates to skill(s)
3. Test that changes improve clarity/correctness
4. Write clear PR description with rationale
5. Submit PR for review
6. Respond to feedback and iterate

Consult references/pr-workflow.md for detailed process.

Contribution Quality Standards

Good Contributions Include

Clear rationale:

"I encountered rate limiting with OpenRouter 5 times. Added exponential
backoff pattern with jitter which resolved all instances. This pattern
isn't documented anywhere in ai/models/."

Before/after comparison:

Before: "Use memory_rethink for updates"
After: "Use memory_insert for concurrent writes (safer), memory_rethink
only for single-agent exclusive access"
Why: Prevents data loss in multi-agent scenarios

Evidence of validation:

"Tested across 3 different projects, pattern held. Also confirmed in
Letta docs. Previous approach caused data loss 2/3 times."

Preserved existing knowledge:

• Don't delete working information
• Build on rather than replace
• Add context, don't remove context

Avoid These Anti-Patterns

❌ Premature optimization - Changing after single instance without validation

❌ Over-generalization - "This worked for me once" → "Always do this"

❌ Opinion as fact - Personal preference without objective improvement

❌ Churn - Changes that are different but not better

❌ Deleting context - Removing information that might help others

Common Pitfalls

Even well-intentioned contributions can miss the mark. Here are patterns to watch for:

Pitfall 1: The Specificity Trap

Pattern: Documenting your specific solution instead of extracting the general pattern.

Example - TOO SPECIFIC:

Skill: git-workflow-manager
Content: "Always end commits with: Written by Cameron ◯ Letta Code"
Problem: This is Cameron's personal preference, not general knowledge

Example - APPROPRIATELY GENERAL:

Skill: git-workflow-manager (revised concept)
Content: "Check repository for commit conventions in CONTRIBUTING.md or recent commits"
Better: Teaches pattern of discovering conventions, applies to any repository

How to avoid:

• Ask: "Would this help an agent on a completely different project?"
• Look for personal names, specific URLs, environment-specific configs
• Extract the pattern, not just your implementation

Pitfall 2: Documentation vs. Specialized Knowledge

Pattern: Creating skills that just reformat well-known documentation.

Example - JUST DOCUMENTATION:

Skill: "How to use git"
Content: Explains git commands, branching, committing, PR creation
Problem: This is standard git knowledge available everywhere

Example - SPECIALIZED KNOWLEDGE:

Skill: "mcp-builder" (in this repository)
Content: Patterns for creating MCP servers - not just "how to use the MCP protocol"
but specific guidance on tool design, error handling patterns, testing strategies
Better: Takes general protocol knowledge and adds specialized patterns for building quality servers

How to avoid:

• If official docs cover it well, link to docs instead of recreating
• Ask: "What specialized insight am I adding beyond standard documentation?"
• Focus on non-obvious patterns, edge cases, or agent-specific considerations

Pitfall 3: Over-Generalizing From One Instance

Pattern: "I made one mistake" → "Let's create 3 new skills to prevent it"

Real example from this repository (November 2025):

Observation: One agent submitted overly-specific git-workflow-manager skill
Initial reaction: "We need CULTURE.md + 3 new skills (knowledge-curation,
agent-human-collaboration, pattern-recognition) + extended documentation"
Correction: Another agent called this out as the exact over-generalization we warn against
Right-sized solution: Add this "Common Pitfalls" section instead
Learning: The system worked - peer review caught premature abstraction

How to avoid:

• Count your data points: 1 occurrence = note it, 3+ = consider contributing
• Check if existing skill can be extended instead of creating new one
• Ask: "Am I solving a real recurring problem or reacting to one experience?"

Pitfall 4: The Abstraction Ladder Confusion

Understanding where your learning sits on the abstraction ladder:

Level 1 - Specific Solution (TOO LOW for skills):

"I configured Firebase Auth with Google OAuth by setting these environment
variables and calling these specific API endpoints"
→ Too specific, only helps others using Firebase + Google OAuth

Level 2 - Pattern (GOOD for skills):

"OAuth integration strategies: Provider discovery, token management, callback
handling, session persistence. Applies to Google, GitHub, Microsoft, etc."
→ General enough to help with any OAuth integration

Level 3 - Principle (ALSO GOOD for skills):

"Delegating authentication to specialized providers: Trade-offs between
managed auth services vs. self-hosted, security considerations, user experience"
→ Helps with authentication decisions broadly

Where to contribute:

• Level 1: Keep in project-specific docs or memory, not skills
• Level 2: Perfect for skills - concrete patterns that generalize
• Level 3: Great for skills - principles that apply across domains

How to climb the ladder:

• Start with: "What did I do?"
• Extract: "What pattern was I following?"
• Distill: "What principle guided this?"
• Contribute at Level 2 or 3

Pitfall 5: Personal Preferences as Best Practices

Pattern: "I like doing it this way" → "Everyone should do it this way"

Example:

"Always use arrow functions in JavaScript"
"Always put API calls in src/api/ directory"
"Always use Claude Sonnet over GPT-4o"
→ These are preferences without objective evidence of superiority

How to avoid:

• Ask: "Can I measure why this is better?" (speed, reliability, cost, etc.)
• Consider: "Are there valid reasons someone would choose differently?"
• Test: "Does this improve outcomes or just match my style?"

Pitfall 6: Fragmenting Information

Pattern: Creating new skills/docs when information should be added to existing ones.

Signs you're fragmenting:

• New skill significantly overlaps with existing skill's domain
• Information could be a section in existing skill
• Creates "which skill do I check?" confusion
• Duplicates concepts across multiple skills

How to avoid:

• Review existing skills in the domain first
• Ask: "Does this extend an existing skill or truly need separate space?"
• Bias toward extending existing skills unless clearly distinct domain

Self-Check Before Contributing

Ask yourself:

1. ❓ Is this general enough? (Would it help agents on different projects?)
2. ❓ Is this specialized enough? (Does it add insight beyond standard docs?)
3. ❓ Is this validated enough? (Have I seen this pattern 2+ times?)
4. ❓ Is this objective enough? (Based on evidence, not preference?)
5. ❓ Is this appropriately placed? (New skill vs. extend existing vs. don't contribute?)

If you can confidently answer yes to all five → Contribute

If you're unsure on any → More validation needed or reconsider contribution type

PR Description Template

Use this template for skill contributions:

## What

[Brief description of what's changing]

## Why

[Problem you encountered / gap you found / improvement opportunity]

## Evidence

[How you validated this is better / how you tested / how often you saw this]

## Impact

[Who benefits / what situations this helps / what it prevents]

## Testing

[How you verified the change works / examples you tried]

See references/contribution-examples.md for real examples.

Example Workflows

Workflow 1: Correcting Existing Skill

During task: "Following letta-memory-architect, I used memory_rethink for
concurrent updates. Result: data loss when two agents wrote simultaneously."

Validation: "Checked references/concurrency.md - it says memory_insert is
safer but warning wasn't prominent. Tested memory_insert with concurrent
writes - no data loss."

Action:
1. Create feature branch: fix/memory-concurrency-warning
2. Update SKILL.md to make warning more prominent
3. Add concrete example of data loss scenario
4. Create PR: "Emphasize memory_insert for concurrent writes"
5. Explain in PR: "Misread the guidance, led to data loss. Making warning
   more visible to prevent this for other agents."

Workflow 2: Adding Missing Pattern

During task: "Hit OpenRouter rate limits 5 times across different projects.
Spent 30min each time figuring out exponential backoff."

Validation: "Pattern works consistently. Checked ai/models/ - not documented.
This is generalizable beyond my specific use case."

Action:
1. Create feature branch: add/api-rate-limiting-skill
2. Create new skill: ai/models/api-rate-limiting/
3. Document exponential backoff pattern with code examples
4. Create PR: "Add API rate limiting patterns"
5. Explain: "Common pattern that caused repeated debugging time. Validated
   across 5 instances with different APIs."

Workflow 3: Clarifying Ambiguity

During task: "Skill said 'use appropriate model' but didn't define criteria.
Tried GPT-4o, Claude Sonnet, GPT-4o-mini before finding best fit."

Validation: "Through testing, identified that task complexity + budget
constraints should guide model choice. This clarification would have saved
1 hour."

Action:
1. Create feature branch: clarify/model-selection-criteria
2. Add decision tree to skill
3. Include examples: "For X task → Y model because Z"
4. Create PR: "Add model selection decision tree"
5. Explain: "Ambiguous guidance led to trial-and-error. Adding decision
   criteria to help agents choose upfront."

Self-Correction Culture

When you make mistakes:

• Note what you learned
• Update relevant skill if gap exists
• Don't just fix the instance, prevent future instances

When you discover better approaches:

• Compare objectively with current documented approach
• Test to validate improvement
• Propose update with clear reasoning

When skills lead you astray:

• Don't assume skill is wrong without investigation
• Validate your alternative approach
• If truly better, propose improvement with evidence

Validation Questions

Before submitting PR, ask:

1. Is this a genuine improvement or just different?
2. Have I validated this works better?
3. Is my evidence strong enough?
4. Am I preserving existing valid knowledge?
5. Will other agents benefit from this?
6. Is my PR description clear about what and why?

Building on Others

Attribution:

• Reference existing skills you're building on
• Credit agents/humans whose insights helped
• Link to forum discussions or sources

Collaboration:

• Respond to PR feedback constructively
• Iterate based on reviewer insights
• Merge after approval, don't force through

Continuous improvement:

• Your contribution will be built upon by others
• This is expected and encouraged
• Living knowledge base means constant evolution

Next Steps

After contributing:

1. Watch for PR feedback and respond
2. Note if your learning helps in future tasks
3. Continue pattern recognition in your work
4. Build on what you contributed as you learn more

The goal: A knowledge base that gets smarter with every agent interaction.