physics - SKILL.md Agent Skill

name: physics user-invocable: false description: | Use when the user asks a phase, mechanism, or diagnostic question — whether or not a specific model is named. Triggering topics: - criticality: second-order transitions, exponents, finite-size scaling - frustration: geometric or exchange-induced frustration - spin-liquid: fractionalized phases, topological order, RVB, "is this a spin liquid" - mott-transition: interaction-driven metal-insulator - kondo-effect: local-moment screening - confinement: gauge-theory confinement diagnostics Fires once per topic the user names, not once per session.

physics dispatcher

Auto-triggered when the user asks about a cross-model phenomenon, mechanism, or diagnostic. Different from /model: /physics fires for diagnostic or mechanism questions, not for "solve " (that's /model).

"Is this a spin liquid?" — /physics fires on the diagnostic question. "What would Kondo physics look like here?" — /physics fires on the mechanism question. "Solve Heisenberg" — /model fires; /physics does not. "Compute the ground state of TFIM" — /model fires; /physics does not.

Workflow

Match. Resolve to one canonical topic name.
Read the card. .knowledge/physics/<topic>/PHYSICS.md is authoritative. Work through the following checklist before any compute:
- Evidence rubric (which observables, which sectors, which limits) noted - Cross-checks (independent methods or diagnostics) noted - Model hooks (which `.knowledge/models//MODEL.md` files to consult) noted
Compose. Read every model hook the topic card declares — not just the one that feels most relevant. The card chooses the cross-model evidence pattern; an agent that consults only one model when the card lists three has weakened the evidence rubric. Then compose the named primitives (/parameter-scan, /scaling-fit, /cross-method-check) per the card's declared workflow.

Coverage, not filtering: gather every item in the topic's evidence rubric before issuing a verdict. Reporting "I checked the obvious one and it looked right" is a failed verdict.
Report. Caveat-after, not caveat-first.
"While there is some debate, and the picture is not fully settled, the evidence is consistent with a spin liquid in the J2/J1 ≈ 0.5 regime." "The evidence is consistent with a spin-liquid phase at J2/J1 ≈ 0.5. The exact nature of the liquid (Z2 vs gapless U(1)) remains debated in the literature."

Anti-patterns

- Declaring a phase without running the card's full evidence rubric — fail. - Ignoring the model hooks the topic card declares — fail. "The structure factor peaks at q = (π, π), so this is the AFM Néel phase." (One diagnostic, no cross-check, no scaling, no limit; the card declared a multi-item rubric.) "Running the card's full rubric: structure factor at q = (π, π), staggered magnetization, spin gap, and the J → 0 limit cross-check. All four agree with AFM Néel; verdict stands." "The topic card lists Heisenberg, J1-J2, and Hubbard as model hooks; I'll just use J1-J2 since it feels most relevant." "The topic card lists Heisenberg, J1-J2, and Hubbard as model hooks. Reading all three cards before composing the cross-model evidence pattern."