---
name: continuous-learning
description: Extract reusable patterns from sessions, score confidence, and
  promote high-confidence instincts into durable skills. Use after completing
  significant work.
metadata:
  category: model-invoked
---

## Pi Context-Aware Execution

When this skill is invoked in Pi, treat the user's current request and any skill arguments as the task input. Do not treat this file as the task by itself.

Before applying the skill, establish only the context needed for the request:

1. Identify the current working directory and relevant project scope.
2. Read local guidance first when present: AGENTS.md, CLAUDE.md, TODO.md, or nearby task/spec files.
3. Inspect the current codebase with targeted searches (prefer rg) and read relevant files before making claims or proposing changes.
4. Ground findings and recommendations in project evidence: cite file paths, commands, tests, docs, or external sources as applicable.
5. Ask a concise clarification only when the arguments and codebase context are insufficient to proceed safely.

Operate conservatively: avoid broad scans, large reads, subagents, or parallel fanout unless the user's requested depth clearly requires them.
Default output: return only the result, blockers, and required evidence. Omit preambles, process narration, repeated context, confidence scores, and follow-up offers. Use at most five bullets unless a required artifact or schema needs more.

# Continuous Learning

## Overview

Learn from sessions in two complementary ways:

1. **Instinct-based learning** — extract reusable patterns, score confidence, promote high-confidence instincts into skills.
2. **Workspace memory** — delegate durable preferences and facts to the `agents-memory-updater` subagent (orchestration only in this skill).

## When to Use

- After completing a significant feature or fix (instincts)
- When a session revealed a non-obvious solution worth remembering
- Before starting similar work to apply learned instincts
- When transcript deltas may update `AGENTS.md` (workspace memory)

## Instinct Lifecycle

### 1. Discovery

When a pattern is encountered during work:

```
Pattern: [What was discovered]
Context: [When this applies]
Confidence: [0.0-1.0 based on evidence]
Evidence: [What proves this works]
```

### 2. Validation

Before storing an instinct:

- [ ] Pattern has been verified in at least one successful outcome
- [ ] Context is specific enough to avoid false positives
- [ ] Confidence score is justified with evidence
- [ ] No conflicting instincts exist

### 3. Storage

Instincts are stored with:

```yaml
instinct:
  name: "descriptive-name"
  pattern: "what to do"
  context: "when to apply"
  confidence: 0.85
  evidence: "why this works"
  created: "2026-01-15"
  usage_count: 3
  success_rate: 0.95
```

### 4. Evolution

When an instinct reaches thresholds:

- Confidence > 0.9 AND usage_count > 5 AND success_rate > 0.9 → Promote to skill
- Confidence < 0.5 AND usage_count > 3 → Deprecate
- Confidence unchanged after 30 days → Review

## Confidence Scoring

| Score | Meaning | Action |
|-------|---------|--------|
| 0.9-1.0 | Proven pattern, multiple successes | Promote to skill candidate |
| 0.7-0.9 | Strong pattern, some evidence | Use with confidence |
| 0.5-0.7 | Plausible pattern, limited evidence | Use cautiously, verify |
| 0.3-0.5 | Weak pattern, speculative | Note but don't rely on |
| 0.0-0.3 | Unproven, likely incorrect | Discard |

## Import and Export

### Export Instincts

Export all instincts to JSON: `[{name, pattern, context, confidence, evidence}]`

### Import Instincts

Import instincts from JSON. Merge with existing records and update confidence on duplicates.

## Workspace memory (AGENTS.md)

When transcript mining may produce durable updates—not one-off task context:

1. Call `agents-memory-updater`; return its result unchanged.
2. Do not mine transcripts or edit files in the parent flow.

The updater owns `AGENTS.md` sections (`## Learned User Preferences`, `## Learned Workspace Facts`), incremental transcript indexes under `~/.cursor/projects/<workspace-slug>/agent-transcripts/`, and deduplication (max 12 bullets per learned section).

If no meaningful updates: respond exactly `No high-signal memory updates.`

## Anti-Rationalization Table

| Excuse | Counter |
|--------|---------|
| "I'll remember this pattern" | Human memory is unreliable. Document it now with context and evidence. |
| "This is too specific to be useful" | Specific patterns become general skills through evolution. Start specific, generalize later. |
| "I don't have time to document" | Two minutes now saves hours of rediscovery later. Use the instinct template. |
| "The confidence score is subjective" | Confidence is a starting point. Usage and success rates provide objective data over time. |