empirica-meta

Empirica Meta-Agent Skill

Philosophy

"Same epistemic rules apply at every meta-layer."

This skill enables recursive self-improvement of Empirica using Empirica's own epistemic framework. When working on Empirica itself, apply the full CASCADE workflow.

When to Use

Activate when:

• Fixing bugs in Empirica CLI or core
• Adding new features to the epistemic framework
• Reviewing/updating architecture documentation
• Investigating why workflows aren't working
• Proposing system improvements

Workflow

1. PREFLIGHT - Assess Current Understanding

Before modifying Empirica, assess what you know about the specific area. Use vectors: know, uncertainty, context, engagement.

2. NOETIC - Investigate Before Acting

• Read relevant architecture docs first
• Search for similar past changes via project-search
• Check for unknowns that might be related
• Log findings as you discover them

3. CHECK - Gate Before Implementation

Run CHECK to verify you're ready to modify the system. Gate: know >= 0.70 AND uncertainty <= 0.35 (after bias correction)

4. PRAXIC - Implement with Care

• Follow self-improvement protocol from CLAUDE.md
• Prefer minimal edits
• Never modify core safety constraints
• Log high-impact findings (0.8+)

5. POSTFLIGHT - Measure Learning

• What did I learn about Empirica's architecture?
• Did the change work as expected?
• Are there follow-up improvements?

Key Files

Self-Improvement Protocol

1. Identify - Recognize gaps through noetic investigation
2. Validate - Test the improvement before proposing
3. Propose - Tell user what you found and suggested fix
4. Implement - If approved, make minimal precise edits
5. Log - Record as finding with impact 0.8+

Turtle Stack

Layer 4: Meta-Orchestrator (future) Layer 3: Sentinel - aggregate, arbitrate, merge Layer 2: Epistemic Agent - spawn, investigate, report Layer 1: CASCADE Workflow - PREFLIGHT/CHECK/POSTFLIGHT Layer 0: Breadcrumb Trail - findings, unknowns, dead ends

Each layer uses same 13 vectors. This skill operates at Layer 2-3.

Gotchas

• Always read before editing (even for Empirica code)
• The system prompt is in ~/.claude/CLAUDE.md (global)
• Skill files go in .claude/skills/ (project)
• Condensed skills go in project_skills/ (for bootstrap)
• Check unknowns before logging new ones
• Commit after each goal (prevent drift)
• Sentinel is now wired via MCP (EMPIRICA_EPISTEMIC_MODE=true)
• PreToolCall hooks gate Edit/Write/Bash via CHECK

References

• empirica --help - Full CLI reference
• docs/architecture/separation-of-concerns.md - What goes where
• docs/architecture/EPISTEMIC_AGENT_ARCHITECTURE.md - Turtle stack
• docs/architecture/QDRANT_EPISTEMIC_INTEGRATION.md - Semantic search