model - SKILL.md Agent Skill

name: model user-invocable: false description: | Use when the user names or describes a harness-tracked quantum lattice model. Match user prose to one of: - transverse-field-ising (TFIM): quantum-critical Ising chain / 2D Wilson-Fisher - heisenberg: SU(2) magnet, AFM or FM by sign of J - j1-j2: frustrated Heisenberg, J2/J1≈0.5 spin-liquid candidate - t-v: spinless fermions + NN repulsion (CDW vs Luttinger) - hubbard: t-U electrons; Mott transition, cuprate parent - t-j: strong-coupling Hubbard with no-double-occupancy - anderson-impurity (SIAM): impurity-in-bath, Kondo - multiorbital-hubbard: multi-band + Hund's J - spin-1-xxz: Haldane phase, AKLT - potts-clock: q-state, first-order / continuous / BKT by q Fires for each named model the user touches in a session, not just the first match.

model dispatcher

Auto-triggered. The user does not type /model; the description above fires the skill when their prose names a harness-tracked model.

Audience definition (binding)

The reader is a working physicist with no harness-internal context. They want the result with embedded reasoning (what method, why, what was verified), not the agent's process. They do NOT know harness vocabulary (manifest, deviation). Every user-facing line is anchored to this audience.

Workflow

Match. Resolve user's prose to one canonical model name. Handle aliases (TFIM → transverse-field-ising, SIAM → anderson-impurity, …).
Read the card. .knowledge/models/<name>/MODEL.md is authoritative; agent memory is not. Work through the following checklist before any compute:
- Hamiltonian definition and sign/normalization conventions read - Declared phases and their order parameters identified - Observables and their canonical forms noted - Recommended method(s) and their stack noted - Verification rubric (limit / symmetry / convergence / cross-method) noted
Execute. Follow the card's declared workflow. The card names which primitive skills to compose (/solve, /parameter-scan, /scaling-fit, /cross-method-check, /using-slurm, /reproduce-paper). Do NOT substitute a different primitive because it feels more familiar; the card's choice is authoritative.
Report. Three lines or fewer in prose: primary quantity (value + units), verification status (which checks passed), and one-line reasoning (method + why). Auto-generate and embed the relevant convergence or stability plot — never report a result without the visual proof it converged.

Anti-patterns

- Substituting generic ED/DMRG defaults for the card's declared workflow — fail. - Acting on agent memory ("I remember Heisenberg has 3 phases") instead of re-reading the card — fail. Memory drifts; cards don't. "Heisenberg has 3 phases — AFM, FM, and PM — so I'll measure ⟨S·S⟩ and ⟨Sᶻ⟩." "Re-reading .knowledge/models/heisenberg/MODEL.md before naming phases. The card declares [list from card]; observables [list from card]."