continuous-learning-v2

name	continuous-learning-v2
version	1.0.0
compatibility	Any AI coding agent (Antigravity, Claude Code, Copilot, Cursor, OpenCode, Codex, pi, and all tools supporting the Agent Skills open standard)
description	Instinct-based learning system that automatically extracts and evolves patterns from AI sessions. Use to build a growing library of project-specific preferences and workflows that improve the AI's effectiveness over time. Covers instinct extraction with confidence weighting, conflict detection, instinct decay, and promotion of confirmed patterns to skills/workflows.
category	meta-learning
triggers	["learn from this session","extract patterns","instinct","continuous learning","remember this pattern","evolve my skills","what have you learned","capture learning"]
dependencies	[{"skill-stocktake":"recommended"},{"autoresearch-loop":"recommended"},{"skill-generator":"recommended"}]

Continuous Learning v2 — Instinct System

Identity

You are a meta-learning specialist. Your job is to extract behavioral patterns from development sessions, store them as atomic "instincts", and evolve them into refined skills, commands, or agents over time. You make the AI system smarter with every session.

Your core responsibility: Continuously improve the AI system by extracting, storing, and promoting learned patterns from every session.

Your operating principle: Every user correction is a learning signal; every repeated pattern is an instinct candidate.

Your quality bar: Every session produces at least one observation with evidence and confidence score; no instinct is added without checking for conflicts; high-confidence instincts are promoted to personal/; expired instincts are retired — no exceptions.

When to Use

• After a session where the user corrected your approach
• When you notice a repeated workflow pattern worth capturing
• When the user says "remember this" or "always do X"
• At the end of long sessions to extract what was learned
• To import/export patterns between projects or team members

When NOT to Use

• Mid-task — do not interrupt active implementation to run the learning loop; save it for end-of-session
• When there is nothing new to extract (no user corrections, no novel patterns, no repeated observations from this session)
• As a substitute for documentation — instincts capture behavioral patterns, not architecture decisions or API contracts
• When the session was purely exploratory with no confirmed outcomes (low-confidence noise is worse than no instinct)

---

Core Principles

1. Atomic instincts. One trigger, one action. Instincts that bundle multiple behaviors cannot be evaluated independently and create conflict resolution complexity.
2. Confidence-weighted. Every instinct has a confidence score (0.3-0.9). Instincts below 0.7 are not applied automatically; they are surfaced for user confirmation.
3. Evidence-backed. Every instinct must reference the specific observations that created it. An instinct without evidence is a preference, not a learned behavior.
4. Conflict-aware. Before adding an instinct, search existing instincts in the same domain for conflicts. Conflicting instincts cause non-deterministic behavior.
5. Decay over time. Unvalidated instincts lose relevance. After 90 days without re-validation, confidence drops. After 365 days, instincts expire.
6. Evolve or retire. Three related instincts in the same domain should become a skill. An instinct that is never validated should be retired.

---

The Instinct Model

An instinct is a small, atomic learned behavior with confidence weighting:

---
id: prefer-functional-style
trigger: "when writing new functions"
confidence: 0.7
domain: "code-style"
source: "session-observation"
scope: project
project: "my-react-app"
---

# 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 2026-03-16

How the Learning Pipeline Works

Session Activity -> Observation Log (.agent/instincts/observations/)
                  -> Raw Instincts (.agent/instincts/personal/)
                  -> Evolved Artifacts (.agent/skills/ or .agent/workflows/)

Blocking Violations (NEVER)

Violation	Consequence	Recovery
Promoting instinct from single observation	One data point treated as universal pattern; no context recorded	Require >= 2 independent observations before promotion
Adding instinct without checking for conflicts	Contradictory instincts cause non-deterministic behavior	Search existing instincts in same domain before adding
Writing observations only at session end	Key decision moments forgotten by summary time	Write observation immediately after each significant decision
Assigning high confidence without validation data	Confidence without evidence is indistinguishable from bias	Only assign high confidence with verifiable evidence
Skipping instinct file format validation	Malformed file fails silently; agent regresses to baseline	Validate YAML/JSONL format after every write

Verification

Self-Verification Checklist

• At least 1 instinct file written per user correction
• Every instinct has non-default confidence score
• Instinct files valid JSONL/YAML
• High-confidence instincts promoted or justified
• Every instinct has non-empty evidence field
• No conflict between new and existing instincts

Verification Commands

# Count observations
ls .agent/instincts/observations/ | wc -l

# Validate instinct files
python -c "import json,yaml; json.load(open('file.jsonl'))"

# Check for conflicts
grep -rn "domain:" .agent/instincts/personal/*.yaml | sort | uniq -c | sort -rn

# Check confidence scores
grep "confidence:" .agent/instincts/personal/*.yaml

Quality Gates

Gate	Criteria	Fail Action
Observation Count	>= 1 per user correction	Capture at least one observation per correction
Evidence Quality	All instincts have non-empty evidence field	Add evidence before saving instinct
Conflict Check	No conflicts with existing instincts	Resolve conflicts before adding
Format Validation	Files parse as valid JSONL/YAML	Fix format errors before session end

Examples

Example 1: Correction Extraction

User request: "No, use Vitest not Jest. I've told you this before."

Skill execution:

1. Create instinct: prefer-vitest.yaml, confidence=0.9
2. Evidence: "User explicitly stated preference for Vitest over Jest"
3. Check for conflicts: no conflicts found
4. Promote to personal/ immediately (high confidence + explicit instruction)

Result: Instinct saved. Future test setup will default to Vitest.

Example 2: Edge Case - Conflict Detection

User request: "Actually, for this monorepo, use Jest because Vitest has issues with the workspace setup."

Skill execution:

1. Compare with existing prefer-vitest.yaml (confidence=0.9, scope=global)
2. Conflict detected: new instinct says Jest, existing says Vitest
3. Resolution: new instinct has supersedes: prefer-vitest, scope=project
4. Old instinct gets superseded_by: prefer-jest-monorepo, confidence reduced

Result: Project-specific override of global instinct. Both coexist with clear hierarchy.

Anti-Patterns

• Never promote an instinct from a single observation because one data point is insufficient to distinguish a genuine pattern from a task-specific anomaly.
• Never add an instinct without checking for conflicts with existing instincts because contradictory instincts cause non-deterministic agent behaviour that is hard to diagnose.
• Never write observations at session end because the key decision moments that produce the most valuable learning happen mid-session and are forgotten by the time the summary is written.
• Never assign high confidence to an instinct without validation data because confidence without evidence is indistinguishable from bias.

Failure Modes

Failure	Cause	Recovery
Instinct extracted from single session over-generalised	One data point treated as universal pattern; context not recorded	Tag every instinct with source session and domain; require 2+ observations
Low-confidence observation promoted without validation	Agent assigns high confidence to pattern observed only once	Only promote to active when confidence >= 0.7 after validation
Instinct conflicts with existing, causing contradictory behaviour	New instinct added without checking for conflicts	Search instinct file for related entries; resolve conflicts with supersedes note
Session observation written too late	Observation captured at end-of-session; key moments forgotten	Write observations immediately after each significant decision

Performance & Cost

Model Selection

Task	Recommended Model	Cost per session
Observation extraction (per session)	Haiku	$0.01-$0.03
Instinct conflict check (10 instincts)	Haiku	$0.02-$0.05
Instinct promotion evaluation	Sonnet	$0.05-$0.10
Skill evolution (3+ instincts → skill)	Sonnet	$0.10-$0.25

Token Budget

• Observation entry: ~200-500 tokens per observation
• Instinct file: ~300-800 tokens per instinct
• Full learning pipeline run: ~2-4KB total
• Expected context usage: 1-3KB per end-of-session pipeline
• When to context-optimize: When processing 20+ observations from a long session

References

Internal Dependencies

• skill-stocktake — Periodic audit of instincts for relevance and overlap
• autoresearch-loop — Uses instinct library as input for broader skill improvements
• skill-generator — Converts 3+ related instincts into a new skill

External Standards

• Spaced Repetition — Inspiration for instinct decay mechanism (90/180/365 day intervals)

Related Skills

• skill-stocktake — Follows continuous-learning-v2 for periodic instinct audit
• autoresearch-loop — Uses extracted instincts as input for system-wide improvements

Changelog

Version	Date	Changes
2.0.0	2026-07-09	Upgraded to Gold Standard v2.0: added frontmatter version/category/dependencies, Core Principles, Blocking Violations table, Verification with commands/quality gates, Examples, References, Changelog.

---