system2-attention

npx skills add majiayu000/claude-skill-registry --skill system2-attention --agent claude-code

Same command for any agent — swap --agent for codex, cursor, copilot.

# System 2 Attention Skill: Deliberate Reasoning Validation **Status**: ✅ Production Ready **Trit**: -1 (MINUS - validator/constraint) **Color**: #2626D8 (Blue) **Principle**: Filter noise via deliberate re-attention **Frame**: Two-stage attention with explicit reasoning --- ## Overview **System 2 Attention** (S2A) validates and filters transformer attention by regenerating context deliberately. Standard attention (System 1) is fast but susceptible to sycophancy and irrelevant context. S2A re-attends after explicit reasoning. 1. **Context regeneration**: LLM rewrites context removing irrelevant info 2. **Two-pass attention**: Fast then deliberate 3. **Sycophancy reduction**: Filter opinion-seeking noise 4. **Factual grounding**: Anchor to verified facts ## Core Pattern ``` S2A(x, context): # System 1: fast pattern matching context_filtered = LLM("Extract only relevant facts from: {context}") # System 2: deliberate reasoning on clean context return LLM(x, context=context_filtered) ``` ```python def system2_attention(query: str, context: str, model) -> str: # Stage 1: Regenerate context (remove sycophantic/irrelevant) filter_prompt = f"""Given the context below, extract only the obje

Apply S2A filtering
just s2a-filter context.txt query.txt
Measure sycophancy reduction
just s2a-sycophancy-test model responses/
Validate factual grounding
just s2a-grounding response.txt facts.txt