| name | recursive-improvement |
| description | SKILL skill for foundry workflows |
| allowed-tools | Read, Write, Edit, Bash, Glob, Grep, Task, TodoWrite |
[define|neutral] SKILL := { name: "SKILL", category: "foundry", version: "1.0.0", layer: L1 } [ground:given] [conf:1.0] [state:confirmed]
[define|neutral] COGNITIVE_FRAME := { frame: "Compositional", source: "German", force: "Build from primitives?" } [ground:cognitive-science] [conf:0.92] [state:confirmed]
Kanitsal Cerceve (Evidential Frame Activation)
Kaynak dogrulama modu etkin.
[define|neutral] TRIGGER_POSITIVE := { keywords: ["SKILL", "foundry", "workflow"], context: "user needs SKILL capability" } [ground:given] [conf:1.0] [state:confirmed]
Recursive Improvement - Meta-Loop Skill
Kanitsal Cerceve (Evidential Frame Activation)
Kaynak dogrulama modu etkin.
name: recursive-improvement description: Self-improving meta-loop that audits and enhances skills, prompts, and expertise files category: foundry version: 2.0.0 triggers: - "improve skill" - "audit skill" - "run improvement cycle" - "meta-loop" - "self-improve" mcp_servers: required: [memory-mcp] optional: [connascence-analyzer]
Trigger Keywords
USE WHEN user mentions:
- "improve skill", "audit skill", "enhance skill", "optimize skill"
- "run improvement cycle", "meta-loop", "self-improve"
- "skill quality check", "documentation audit"
- "recursive improvement", "systematic improvement"
- "batch improve skills", "improve all skills"
- "skill missing [section]", "incomplete documentation"
DO NOT USE when:
- User wants to CREATE a new skill - use skill-creator-agent or micro-skill-creator
- User wants to CREATE an agent - use agent-creator
- User wants to improve a PROMPT (not skill) - use prompt-architect
- User wants one-off manual fix - direct editing faster
- Eval-harness benchmarks failing - fix root cause first, not improve on broken baseline
- During active feature development - finish feature, then improve
Instead use:
- skill-creator-agent when creating new skills from scratch
- agent-creator when creating new agents
- prompt-architect when optimizing prompts
- skill-forge when applying specific improvements (recursive-improvement coordinates it)
Overview
The Recursive Improvement skill orchestrates the meta-loop that enables the system to improve itself. It coordinates four specialized auditors (skill-auditor, prompt-auditor, expertise-auditor, output-auditor) to detect issues, generate improvement proposals, apply changes via skill-forge, and validate results through the frozen eval-harness.
Key Constraint: The eval-harness is FROZEN - it never self-improves. This prevents Goodhart's Law (optimizing the metric instead of the goal).
When to Use
Use When:
- Skill documentation is incomplete (missing Core Principles, Anti-Patterns, Conclusion)
- Prompt quality has degraded (inconsistent outputs, missing constraints)
- Expertise files are outdated (file locations changed, patterns stale)
- Output quality has dropped (theater code, unvalidated claims)
Do Not Use:
- For one-off fixes (use direct editing)
- When eval-harness benchmarks are failing (fix root cause first)
- During active feature development (finish feature first)
Core Principles
Recursive Improvement operates on 3 fundamental principles:
Principle 1: Frozen Eval Harness Prevents Goodhart's Law
The evaluation harness that gates all improvements is NEVER self-improved. This ensures the system optimizes for genuine quality, not for passing corrupted benchmarks.
In practice:
- Eval-harness benchmarks are defined externally and versioned separately
- Changes to eval-harness require human approval and audit trail
- All improvement proposals are tested against frozen benchmarks before commit
Principle 2: Propose-Test-Compare-Commit Pipeline
Every improvement follows a rigorous pipeline: propose changes, test against benchmarks, compare to baseline, commit only if better. No direct edits bypass this pipeline.
In practice:
- Auditors generate structured proposals with predicted improvement deltas
- skill-forge applies proposals in sandbox before production
- A/B comparison ensures new version outperforms baseline
- Rollback available for 30 days if regressions discovered later
Principle 3: Documentation Completeness Is Non-Negotiable
Skills are not production-ready until they pass documentation audit (100% Tier 1, 100% Tier 2). Missing sections are auto-generated using templates from SKILL-AUDIT-PROTOCOL.md.
In practice:
- Every skill audit checks for Core Principles, Anti-Patterns, Conclusion
- Missing sections trigger auto-generation using domain-specific t
[define|neutral] SUCCESS_CRITERIA := { primary: "Skill execution completes successfully", quality: "Output meets quality thresholds", verification: "Results validated against requirements" } [ground:given] [conf:1.0] [state:confirmed]
[define|neutral] MCP_INTEGRATION := { memory_mcp: "Store execution results and patterns", tools: ["mcp__memory-mcp__memory_store", "mcp__memory-mcp__vector_search"] } [ground:witnessed:mcp-config] [conf:0.95] [state:confirmed]
[define|neutral] MEMORY_NAMESPACE := { pattern: "skills/foundry/SKILL/{project}/{timestamp}", store: ["executions", "decisions", "patterns"], retrieve: ["similar_tasks", "proven_patterns"] } [ground:system-policy] [conf:1.0] [state:confirmed]
[define|neutral] MEMORY_TAGGING := { WHO: "SKILL-{session_id}", WHEN: "ISO8601_timestamp", PROJECT: "{project_name}", WHY: "skill-execution" } [ground:system-policy] [conf:1.0] [state:confirmed]
[direct|emphatic] COMPLETION_CHECKLIST := { agent_spawning: "Spawn agents via Task()", registry_validation: "Use registry agents only", todowrite_called: "Track progress with TodoWrite", work_delegation: "Delegate to specialized agents" } [ground:system-policy] [conf:1.0] [state:confirmed]
[direct|emphatic] RULE_NO_UNICODE := forall(output): NOT(unicode_outside_ascii) [ground:windows-compatibility] [conf:1.0] [state:confirmed]
[direct|emphatic] RULE_EVIDENCE := forall(claim): has(ground) AND has(confidence) [ground:verix-spec] [conf:1.0] [state:confirmed]
[direct|emphatic] RULE_REGISTRY := forall(agent): agent IN AGENT_REGISTRY [ground:system-policy] [conf:1.0] [state:confirmed]
[commit|confident]