ProofSkill

ProofSkill is an evidence-based product strategy assessment. Instead of asking learners to explain what they know, it asks them to make a coherent plan, changes a material condition, and measures how well they adapt.

What it does

The Build Week MVP delivers one complete E-commerce / Intermediate journey:

1. A learner signs in and builds an eight-part strategy through guided decision cards.
2. GPT-5.6 Sol generates an adaptive constraint affecting two to four fields.
3. The learner revises each affected decision and makes the adaptation explicit.
4. A final critical decision forces a visible trade-off, action, and guardrail.
5. GPT-5.6 produces a structured seven-competency evaluation.
6. Deterministic code verifies every positive quote against the saved work, applies transparent scoring rules, and persists the private report.

The result is more than a quiz score: it is a reopenable evidence trail showing the learner's original decision, adaptation, contradictions, strongest proof, primary gap, and next challenge. The private dashboard also restores unfinished sessions.

The public demo is precomputed and explicitly labeled. Live assessments require an account. Supabase Auth and Row Level Security isolate each user's sessions, evaluations, evidence, and AI-run metadata.

Technological implementation

• gpt-5.6-sol through the OpenAI Responses API.
• Structured Outputs with Zod for both constraints and evaluations.
• Low reasoning effort for constraint generation and medium for evaluation.
• store: false, privacy-preserving safety identifiers, differentiated failures, and one controlled citation retry.
• Next.js 16 request-scoped auth, protected server routes, idempotent state transitions, and server-only mutations.
• Supabase RLS plus SELECT-only authenticated grants.
• Deterministic evidence validation, versioned scoring, visible caps, and persisted run metadata.

GPT-5.6 proposes the adaptive constraint and evidence draft. It does not own the final report: deterministic application code verifies quotes, applies rubric weights and caps, and prevents unsupported positive evidence from appearing.

Design and impact

ProofSkill turns assessment into an interactive pressure test. Guided cards reduce blank-page anxiety while the adaptive constraint still requires genuine reasoning. The report translates the experience into actionable learning: what the learner proved, where the strategy contradicted itself, and what challenge should come next.

The same pattern can later support product management education, hiring simulations, cohort coaching, and other domains where observable judgment matters more than memorized terminology.

How Codex was used

Codex was the primary Build Week development partner. It helped convert the brief into vertical slices, implement the Next.js and Supabase architecture, design the state machine and least-privilege policies, build the guided interactions, create tests, diagnose production deployments, verify the OpenAI integration, and prepare the final submission package. Human control remained over product choices, credentials, external accounts, live testing, and submission.

Built with

OpenAI GPT-5.6 Sol, OpenAI Responses API, Structured Outputs, Codex, Next.js 16, React 19.2, TypeScript, Supabase Auth, PostgreSQL, Row Level Security, Vercel, Tailwind CSS 4, shadcn/ui, Zod, Vitest, and Playwright.

Built With

• codex
• next.js-16
• openai-gpt-5.6-sol
• openai-responses-api
• playwright
• postgresql
• react-19.2
• row-level-security
• shadcn/ui
• structured-outputs
• supabase-auth
• tailwind-css-4
• typescript
• vercel
• vitest
• zod