Continuous Learning

Continuous Learning for GitHub Copilot CLI

When this skill is invoked, ask the user which action to perform:

Alternatively, the user may directly state what they want (e.g., "extract learnings from this session", "show my instincts"). Match their intent to the appropriate action below.

Action: learn

Analyze the current session and extract patterns worth saving as instincts.

What to Look For

1. Error Resolution (error_resolution) - What error occurred? Root cause? Fix? Reusable?
2. User Corrections (user_corrections) - Did the user correct AI behavior? What was preferred?
3. Debugging Techniques (debugging_techniques) - Non-obvious steps, tool combinations
4. Workarounds (workarounds) - Library quirks, API limitations, version-specific fixes
5. Project Conventions (project_specific) - Codebase conventions, architecture decisions
6. Tool Preferences (tool_preferences) - Preferred tools, search-before-edit workflows
7. Repeated Workflows (repeated_workflows) - Multi-step sequences used multiple times

When to Suggest Learning

After complex work, proactively suggest: "This session had some interesting patterns. Want me to run /continuous-learning to save them?"

Do NOT extract:

• Simple typo fixes
• One-time issues (API outages)
• Trivial syntax errors

Confidence Assignment

• 1 observation: 0.3 (tentative)
• 2-3 observations: 0.5 (moderate)
• User explicitly stated preference: 0.7 (strong)
• Repeated pattern across session: 0.7 (strong)
• Multiple confirming observations: 0.85 (very strong)

Execution Steps

1. Review the current session for extractable patterns
2. For each pattern, determine: id, trigger, action, confidence, domain
3. Check existing instincts to avoid duplicates (run node scripts/copilot/instinct-cli.js status)
4. Present findings to user and ask for confirmation
5. For each confirmed instinct, create a .yaml file in the instincts directory

Storage path (create directory if missing):

• Windows: %USERPROFILE%\.copilot\learned\instincts\personal\
• Unix/macOS: ~/.copilot/learned/instincts/personal/

File format -- save as <id>.yaml:

---
id: prefer-functional-style
trigger: "when writing new functions"
confidence: 0.7
domain: "code-style"
source: "session-observation"
created: "2025-01-22"
last_validated: "2025-01-22"
observations: 5
---

# Prefer Functional Style

## Action
Use functional patterns over classes when appropriate.

## Evidence
- Observed 5 instances of functional pattern preference
- User corrected class-based approach to functional on 2025-01-22

## When to Apply
When creating new modules, prefer function-based patterns with closures over class-based designs.

Update Rules

If an instinct with the same id already exists:

• Increment observations count
• Update last_validated to today
• Increase confidence by 0.05 (cap at 0.95)
• Append new evidence

Action: status

Show all learned instincts grouped by domain with confidence bars.

Execution

Run the CLI tool:

node scripts/copilot/instinct-cli.js status

Or if you prefer a specific domain:

node scripts/copilot/instinct-cli.js status --domain code-style

If the CLI is not available, manually read all .yaml files from:

• ~/.copilot/learned/instincts/personal/
• ~/.copilot/learned/instincts/inherited/

Parse each file and display grouped by domain, sorted by confidence (highest first).

Action: evolve

Cluster 3+ related instincts into higher-level structures.

Execution

Run the CLI tool:

node scripts/copilot/instinct-cli.js evolve

Or manually analyze instincts and determine evolution type:

-> Skill (auto-triggered behaviors): When instincts describe code style, error handling, or patterns that should always apply.

• Example: prefer-functional + use-immutable + avoid-classes -> functional-patterns skill

-> Prompt (user-invoked actions): When instincts describe repeatable user-requested workflows.

• Example: new-table-migration + update-schema + regenerate-types -> /new-table prompt

-> Agent (complex multi-step): When instincts describe deep, multi-step processes.

• Example: debug-check-logs + debug-isolate + debug-reproduce -> debugger agent

Output Locations

Save evolved structures to ~/.copilot/learned/evolved/:

• Skills: evolved/skills/<name>/SKILL.md
• Prompts: evolved/prompts/<name>.prompt.md
• Agents: evolved/agents/<name>.agent.md

描述撰寫指引（關鍵）

演化產生的每個結構，其 description 欄位必須包含觸發情境與關鍵字，否則 AI 不會在正確時機載入它。

差的描述（缺少觸發情境）：

description: Enforce functional programming patterns

好的描述（包含觸發情境與關鍵字）：

description: 撰寫新函式、建立模組、設計資料流時套用函數式程式設計模式。涵蓋不可變性、純函式、避免類別繼承等慣例。

規則：

1. 開頭用「當...時」或動詞片語描述何時該載入此結構
2. 包含使用者可能在提示中使用的關鍵字
3. 遵循專案語言設定（若有）撰寫描述
4. 長度建議 40-100 字，避免過短（觸發不到）或過長（噪音太多）

Action: export

Export instincts to a shareable YAML file.

Execution

node scripts/copilot/instinct-cli.js export --output instincts-export.yaml

Or with filters:

node scripts/copilot/instinct-cli.js export --domain testing --min-confidence 0.7

Exported files contain only patterns (no code, no file paths, no session data).

Action: import

Import instincts from a YAML file.

Execution

node scripts/copilot/instinct-cli.js import path/to/instincts.yaml

Preview first with --dry-run:

node scripts/copilot/instinct-cli.js import path/to/instincts.yaml --dry-run

Merge strategy:

• New instincts: added to inherited/ directory
• Duplicates: keep the one with higher confidence
• Conflicts: skip and flag for manual resolution

Instinct Properties

• Atomic -- one trigger, one action
• Confidence-weighted -- 0.3 = tentative, 0.9 = near certain
• Domain-tagged -- code-style, testing, git, debugging, workflow, api-design, security, etc.
• Evidence-backed -- tracks what observations created it

Confidence Scoring

Confidence increases when pattern is repeatedly observed or confirmed. Confidence decreases when user corrects the behavior or it goes unused.

Compatibility

Instincts use the same YAML format as the Claude Code version (stored in ~/.claude/homunculus/), enabling cross-tool sharing between Claude Code and Copilot users.