Skill Evolution

Skill Evolution

Framework for treating skills as living entities that evolve based on execution feedback. Inspired by HKUDS/OpenSpace's self-evolving skill engine.

How to use

• /skill-evolution Review all skills for evolution candidates based on recent execution patterns.

• /skill-evolution <skill-name> Analyze a specific skill's health metrics and propose evolution if needed.

When to use

• A skill consistently fails or produces poor results
• After significant codebase or tool changes that may degrade existing skills
• During periodic maintenance (weekly recommended)
• When the autoresearch loop identifies a skill as underperforming

When NOT to use

• For brand new skills (let them accumulate execution data first)
• During active development of a skill (wait for it to stabilize)

Core Concept: Skills as Living Entities

Static skills degrade over time as:

• Underlying tools change or deprecate
• Codebase patterns evolve
• User needs shift
• Models improve (old constraints become unnecessary)

Solution: Track execution metrics per skill and evolve them through three triggers.

Three Evolution Triggers

Trigger 1: Post-Execution Analysis (Reactive, Per-Task)

After each significant skill application:

1. Did the skill produce the expected output?
2. Were there manual corrections needed?
3. Was the skill's guidance followed or overridden?

Action: If corrections were needed 3+ times, trigger FIX evolution.

Trigger 2: Degradation Detection (Reactive, Cross-Task)

Monitor across multiple executions:

• Success rate dropping below threshold (< 70%)
• Increasing override frequency
• Growing number of exceptions/workarounds

Action: If degradation detected, trigger DERIVED evolution (create improved variant).

Trigger 3: Periodic Health Review (Proactive)

On schedule (weekly/monthly):

• Scan all skills for staleness (no use in 30+ days)
• Check for skills that reference deprecated tools or patterns
• Identify skills with low confidence scores in memory

Action: Flag for review, propose updates or retirement.

Three Evolution Modes

FIX: In-Place Repair

• Minimal targeted changes to fix specific failures
• Preserves original intent and structure
• Applied when root cause is identified and isolated

DERIVED: Enhanced Variant

• Creates improved version from parent skill
• Inherits base structure, adds improvements
• Used when the skill concept is sound but execution needs rethinking

CAPTURED: New Pattern Extraction

• Extracts novel patterns from successful ad-hoc executions
• Creates new skills from recurring successful behaviors
• Maps to Operating Framework promotion policy (3+ successes = template)

Skill Health Metrics

Track these per skill (store in memory or dedicated log):

Evolution Process

1. DETECT: Trigger fires (post-exec, degradation, or periodic)
2. DIAGNOSE: Identify what's failing and why
3. PROPOSE: Generate minimal diff (FIX), new variant (DERIVED), or new skill (CAPTURED)
4. REVIEW: Human approval required before applying
5. APPLY: Update SKILL.md
6. VERIFY: Test evolved skill on the original failure case
7. LOG: Record evolution in memory (date, trigger, changes, outcome)

Integration with Existing Setup

Cross-references

• autoresearch skill: Karpathy-inspired iterative improvement (broader scope)
• prompt-reliability-engine skill: Mode 9 (Self-Improving Skill) for prompt-specific evolution
• skill-creator skill: initial skill creation
• operating-framework skill: promotion policy and maintenance cadence
• OpenSpace (HKUDS): original research on self-evolving AI agent skills