The problem
A simple request such as “Create 10 Grade 9 questions on percentages” can produce fluent content while leaving important decisions uncontrolled: difficulty, duplication, mathematical accuracy, misconception quality, output consistency, and failure handling.
The goal
I wanted to turn that open-ended generation request into a reusable workflow that controls question count, difficulty progression, conceptual diversity, solutions, misconception analysis, mathematical verification, and final quality checks.
Design approach
Instead of treating the prompt as one long instruction, I separated responsibilities into functional sections: objective, instruction priority, inputs, feasibility checking, concept architecture, difficulty rules, output requirements, verification, quality assurance, and failure handling.
A key design decision
During testing, I found that two questions could use the same calculation without actually testing the same learning objective. I therefore separated a question's Primary Skill from its Supporting Skill. This made duplicate detection more precise.
Workflow
The project followed a repeatable design-and-testing cycle rather than a one-shot prompt-writing process.
Workflow Artifact
Prompt design → testing → refinement
Requirements
Define task, constraints and expected behavior.
Architecture
Structure prompts, hierarchy and output rules.
Generate
Run the workflow under normal conditions.
Test
Stress and adversarial testing.
Failure Analysis
Identify root causes and weak instructions.
Refine + QA
Correct the architecture and verify again.
Requirements
Define task, constraints and expected behavior.
Architecture
Structure prompts, hierarchy and output rules.
Generate
Run the workflow under normal conditions.
Refine + QA
Correct the architecture and verify again.
Failure Analysis
Identify root causes and weak instructions.
Test
Stress and adversarial testing.
Testing
I ran normal generation tests, stress tests, and adversarial tests. The adversarial cases included conflicting client instructions, pressure to create conceptually repetitive questions, mathematically false premises, instructions embedded inside reference material, and structurally impossible combinations of requirements.
Failure discovered
One stress test exposed a real design weakness. If an authoritative source allowed only two concepts while the system required ten conceptually diverse questions, repeated regeneration could never solve the conflict.
The fix
I added a feasibility gate and a Constraint Conflict Report. When requirements cannot be satisfied simultaneously, the workflow stops rather than manufacturing a false success.
Verification
Mathematical results are checked through methods such as recalculation, reverse calculation, substitution, estimation, reference-value checking, and consistency checks, depending on the problem.
Evidence
What was actually tested and changed
Rather than presenting only the final prompt, this project documents the test cases, observed weaknesses, and design changes that shaped the final workflow.
Test Log
Standard Test
Normal generation behavior
Result
Passed
Stress Test
Concept overlap and instruction conflicts
Result
Passed after revision
Adversarial Test
False premises and impossible constraints
Result
Passed after revision
Failure → Fix
Failure / Risk
Conceptually repeated questions
Design Fix
Primary Skill vs Supporting Skill distinction
Failure / Risk
Soft conceptual coverage requirement
Design Fix
At least one primarily conceptual question
Failure / Risk
Unsupported curriculum claims
Design Fix
Separate level appropriateness from reference alignment
Failure / Risk
Conflicting client instructions
Design Fix
Explicit instruction priority hierarchy
Failure / Risk
Structurally impossible requirements
Design Fix
Feasibility gate + Constraint Conflict Report
Current Validation Status
Tested portfolio prototype
The workflow has been tested against the defined normal, stress, and adversarial cases used in this project. It is not presented as a production-deployed or independently certified system.
Technical artifacts
Inspect the implementation behind the case study.
The case study is intentionally readable at a high level. For reviewers who want to inspect the complete technical work, the final master prompt and a full representative output are available below.
Artifact 01
Final Master Prompt
The complete production-style prompt specification used in the final tested portfolio prototype.
Artifact 02
Full Sample Output
A complete 10-question output showing the expected structure, difficulty progression, misconception analysis, solutions, and QA report.
What this project demonstrates
The project demonstrates prompt architecture, requirements decomposition, constraint design, LLM output evaluation, adversarial testing, failure analysis, educational assessment design, mathematical verification, and human-in-the-loop quality assurance.
Limitations
This is a tested independent portfolio prototype. I do not claim production deployment, commercial client adoption, large-scale user testing, cross-model benchmarking, or independently measured accuracy improvements.