Skip to main content
The repo ships skill specs that let an AI coding assistant (Claude Code, Cursor, GitHub Copilot Chat, Amazon Q, or any IDE assistant with repo context) add new LeetCode problems end to end. You describe the problem in one sentence; the assistant scrapes, templates, generates, implements the optimal solution, and verifies. Human intervention stays minimal: the prompt, and a review of the result.

​
Set up the context

For best results, have these files in the assistant’s context:

​
Ask for a problem

Example prompt requesting a new problem from the assistant Prompts that work:
Include the problem number and title; adding a tag puts it in a collection.

​
What the assistant does

  1. Scrapes the problem data from LeetCode
  2. Transforms it into the JSON template format, images included
  3. Creates json/problems/{problem_name}.json
  4. Updates tags.json5 with the requested tags
  5. Generates leetcode/{problem_name}/ with the full problem structure
  6. Runs the lint checks, iterating from step 3 until everything passes
Source control view of files created during problem generation Step 5 produces real content under leetcode/{problem_name}/: the full six-file layout with the README description, helpers, playground, and tests, all from the JSON template. The generated solution.py arrives as a typed stub with a TODO, and test_solution.py ships parametrized cases: Generated solution.py with TODO placeholder and type hints Generated test_solution.py with parametrized test cases

​
Implement and verify

The assistant then implements the optimal solution in solution.py (one Solution class) and runs quality assurance per the test QA skill: tests pass, at least 12 cases, and the suite reproduces after a regenerate and restore cycle.
Generated test cases are a starting point, not gospel. The assistant verifies expected values, and so should you before trusting a suite.

​
Batch creation

/batch-problem-creation [count] (default 5) loops the whole workflow over the next problems in the roadmap lists:
The picker skips known unscrapable problems (premium, API issues), which live in .claude/.dev/problem_lists/unscrapable.py. Failed problems are logged and the batch continues; you get a summary with success rate and retry candidates at the end.

​
The pipeline, for reference

Everything above drives this machinery:
A problem template JSON carries the metadata (problem_name, problem_number, problem_title, difficulty, topics, tags) plus the rendered content for each output file: the README description, examples and constraints, helper function sources, solution skeleton, test content, and playground cells. Bulk generation for practice is covered in Collections.

​
Fix drift

leetcode/ is never edited by hand. When bake check-consistency (or CI) reports that leetcode/ no longer matches the templates, edit the JSON template, then regenerate:
To rebuild everything from scratch, bake gen-all-problems deletes leetcode/ and regenerates all problems (it asks first outside CI). Two more skills automate the common fixes: consistency-fix (drift repair) and update-tags (collection membership changes).