generate-problem

Provide parameters like domain, topic, difficulty rating, and Bloom level, then run the pipeline. It dispatches six specialized subagents covering blueprint design, statement writing, solution verification, test generation, quality review, and editorial writing, producing a final_problem.json package.

The final_problem.json includes metadata with difficulty and tags, the full problem statement, a verified solution with correctness proof and complexity analysis, a test suite of 10+ categorized cases, an editorial with progressive hints, and a quality report with review scores.

The Solution Engineer stage outputs a SOLVABILITY_FAILURE verdict when constraints are contradictory or ambiguous. Gate 1 then retries the Problem Writer stage up to twice, then the Architect stage once, before aborting with partial output.

The Quality Reviewer scores 10 Shield criteria from 0 to 10, including learning objective clarity, difficulty calibration, and test coverage. Any score below 8 forces a REVISION verdict, and Sword attacks probe for ambiguity, untested edge cases, and incorrect passing solutions.

Yes. Progress is tracked in a durable ledger file at .superpowers/sdd/progress.md. On restart, the Skill reads the ledger and resumes at the first incomplete stage rather than starting over.

Avoid it when you only need a single artifact, such as just a statement or just test cases, since the pipeline always runs all stages. For targeted work, invoke the individual stage skills like write-statement or generate-tests directly.