Skip to content

CLAUDE.md

Latest commit

 

History

History
104 lines (76 loc) · 3.9 KB

CLAUDE.md

File metadata and controls

104 lines (76 loc) · 3.9 KB

CLAUDE.md — MatlabGraderApp

Deployment

After any code change, always run both steps:

npm run deploy   # builds and pushes to gh-pages (updates the live site)
git push origin main  # pushes source to main

The live app is at https://veriqai.github.io/MatlabGraderProblemGenerator/. npm run deploy is required for users to see changes — pushing to main alone does not update the site.

Architecture

Single-page React app (Vite + TypeScript + Tailwind). No backend — the Anthropic API is called directly from the browser using the user's own API key.

File Role
types.ts All shared types: ProblemType, ClassAssessment, AppState, ProblemOption, Artifacts
api.ts callClaude(), parseJsonResponse(), downloadText(), downloadZip(), toSnakeCase()
prompts.ts All five prompt-builder functions (one per artifact)
App.tsx State machine and 4-stage generation pipeline
components/Stage0Input.tsx Problem type / assessment picker UI
components/Stage1Options.tsx Generated option cards, select/deselect
components/Stage2Generate.tsx Progress + artifact review tabs
components/Stage3Done.tsx Final download page
components/Common.tsx Shared UI primitives — ProblemTypeBadge has a colour per type

Generation pipeline

Each problem goes through 4 sequential Claude calls (steps visible in App.tsx):

  1. Description — buildDescriptionPrompt → *_description.txt
  2. Solution — buildSolutionPrompt → solution.m
  3. Template — buildTemplatePrompt(solution) → template.m
  4. Tests — buildTestsPrompt(solution) → all_tests.m

Problem types

Type What the student submits Key detail
Script .m script Variables assessed in workspace
Function .m function Inputs/outputs assessed directly
Class classdef ClassName.m Assessed via instantiation
Object usage .m script Provided supporting class; student writes script only

Class assessments (selected in Stage 0 when problem type = Class)

Assessment What gets blanked
Constructor — property assignment obj.prop = arg lines in constructor
Constructor — computed property Derived property line only
Instance method Method body
Constant property Value inside properties (Constant) block
Operator overloading Overloaded operator method body

Object usage — solution.m format

solution.m contains two sections separated by %%% delimiters:

%%% SUPPORTING FILE: ClassName.m %%%
classdef ClassName
...
end
%%% STUDENT SCRIPT SOLUTION %%%
...script code...

The instructor pastes the class portion into MATLAB Grader → Supporting Files, and the script portion into Reference Solution. template.m contains the script portion only.

all_tests.m quality rules

These rules are enforced via prompt language in buildTestsPrompt. Every generated test file must:

  • Use %% Test N: description section headers only — no === TEST N === banners.
  • Contain exactly one assessVariableEqual('expr', value) per section.
  • Use 1–3 lines of setup max per section.
  • Use randperm(19)-10 when swap/transposition detection matters; randi([lo,hi]) otherwise.
  • Include at least one hardcoding-detection test (different numeric range from Test 1).
  • No try/catch, fprintf, if/else, whos, dir, script_ran flags, or markdown fences.
  • 3–5 tests maximum.

Adding a new problem type or Class assessment

  1. types.ts — add to ProblemType or ClassAssessment union.
  2. components/Stage0Input.tsx — add radio option (new type) or table row (new assessment).
  3. components/Common.tsx — add a colour entry to ProblemTypeBadge.styles if new type.
  4. prompts.ts — add a branch in each of the five build functions.
  5. Run npm run deploy && git push origin main.