Skip to content

CLAUDE.md

Latest commit

 

History

History
59 lines (43 loc) · 4.51 KB

CLAUDE.md

File metadata and controls

59 lines (43 loc) · 4.51 KB
Raw

PMSM Claude Skills — Workspace Guide

This repository is a Claude Code workspace for building MATLAB/Simulink models of PMSM control (FOC, FCS-MPC, DTC, SMC) and for distilling reference models into reusable skills under anti-contamination discipline.

One-time setup

Before doing modeling work in this workspace:

  1. Install the MATLAB MCP server so MATLAB runs directly instead of through Bash:
    • Source: https://github.com/matlab/matlab-mcp-core-server (MathWorks, MIT)
    • Build per its README, then register it: claude mcp add matlab --scope user -- <path-to-binary>
    • It provides detect_matlab_toolboxes, check_matlab_code, evaluate_matlab_code, run_matlab_file.
  2. Install the MathWorks official MATLAB skills for Claude Code per MathWorks' instructions.
  3. Confirm MATLAB R2024b or later with Simulink, Simscape, Simscape Electrical, and Control System Toolbox.

If the MCP server is not available, run MATLAB through its CLI: matlab -batch "run('script.m')" — and use a background process for long simulations.

Skills

.claude/skills/ is registered automatically when Claude Code opens this folder:

  • motor-pmsm-base — plant + dq conventions + building-blocks SOP + broken-FOC defense. Base for the method skills.
  • motor-fcs-mpc — single-vector Finite-Control-Set MPC current loop.
  • motor-fcs-mpc-dualvector — dual-vector (two-vectors-per-period) FCS-MPC current loop with q-axis deadbeat time allocation. Layers on motor-fcs-mpc.
  • motor-fcs-mpc-trivector — three-vector (two adjacent active + one zero per period) FCS-MPC current loop; a 2×2 dual-axis deadbeat solve nulls both i_d and i_q in one period.
  • motor-dtc-pmsm — Direct Torque Control, αβ frame, Sutikno 6-state switching table.
  • motor-smc-pmsm — Sliding Mode Control speed loop (PD-type sliding + super-twisting) over a dq PI current loop.
  • motor-mfpcc-eso — model-free deadbeat predictive current control (ultralocal model + Extended State Observer + SVPWM); parameter-free current law (only α, ω0). Continuous control set, not FCS-MPC. Layers on motor-pmsm-base.
  • simulink-layout-tidy — general-purpose Simulink layout tidier (compact, overlap-free, honest crossing report). Not motor-specific; supports the layout gate.

Each method skill layers on motor-pmsm-base (motor-fcs-mpc-dualvector additionally builds on motor-fcs-mpc). Read the relevant SKILL.md and its references/ before building.

Methodology

To learn an external reference model and distil it into a new skill, follow workflow/reference_model_learning_workflow.md — the 11-phase Reference Model Learning Workflow. Its non-negotiable disciplines:

  • Learn / build separation — understanding a model and rebuilding it are verified as two separate capabilities.
  • No peeking on rebuild — the rebuild subagent must not read the reference model; otherwise the result collapses into a copy and the skill is hollow.
  • Anti-contamination isolation — give every verification subagent an explicit must-NOT-read list (confidential sources, prior-pass artifacts, methodology internals), and audit its output (grep / mathematical consistency / structural consistency) before trusting it.
  • Theory first — derive the plant equations and control-law formulas before reading any .slx; a model that contradicts the derived formulas is a finding, not a hallucination.

Working rules

  • Plan before building. Write a numbered plan (parameter table, design decisions, build-script structure) and get user approval before the first add_block.
  • shared/ is read-only. Reuse shared/formulas/ and shared/building_blocks/ by reference; do not duplicate or re-derive. Build your models in workspace/.
  • One-click reproducibility. A built .slx must run from a double-click in a fresh MATLAB session — inject all parameters via set_param(mdl, 'InitFcn', ...).
  • Visual check before metrics. Before trusting numerical scores, confirm the motor rotates, iq tracks its reference, the abc currents are AC sinusoids (not DC-locked), and the torque energy balance holds. A failed visual check means a broken implementation regardless of the numbers.

How to work here

  • On failure, change approach — don't repeat a method that just failed; after ~3 distinct attempts, stop and report rather than thrashing.
  • Never swallow an error silently.
  • On long build scripts, re-read your plan every few steps.
  • Be direct: if the user's approach has a flaw, give the counter-argument with reasoning rather than going along with it.