AI Investing Trainer — Beta
A daily training tool that helps beginner investors build real decision-making intuition through reps, AI feedback, and structured reflection. Not a course, not a simulator — a training + feedback system.
Train like an investor in 10 minutes.
The core loop
- Decide. Read a real-world stock scenario (metrics + headlines + signal). Choose Buy / Hold / Pass, set your confidence, and write 2–4 sentences of reasoning.
- Get graded. A senior buy-side analyst persona (powered by Claude or GPT) scores your reasoning and your decision on 0–100, flags what you did well, what you missed, and gives one concrete tip.
- See the outcome. A predefined forward return reveals whether the call was right, wrong, or ambiguous — plus a "luck vs skill" callout when confidence and outcome diverge.
- Reflect. Three structured prompts go into your journal, linked back to the decision and its outcome.
- Track progress. Rolling skill score, decision accuracy, and confidence calibration tell you whether your judgment is actually improving.
File structure
Investing App/
├─ app/
│ ├─ layout.tsx # Global shell + nav
│ ├─ page.tsx # Landing
│ ├─ globals.css # Tailwind + transitions + progress bar
│ ├─ challenge/page.tsx # The full daily flow (state machine)
│ ├─ journal/page.tsx # Past reflections
│ ├─ progress/page.tsx # Skill score, accuracy, calibration, trend
│ ├─ history/page.tsx # [BONUS] Review past decisions, expandable
│ └─ api/evaluate/route.ts # AI evaluation endpoint (Anthropic or OpenAI)
├─ components/
│ ├─ Nav.tsx
│ ├─ ScenarioCard.tsx
│ ├─ DecisionInput.tsx # Buy/Hold/Pass + confidence slider + reasoning
│ ├─ FeedbackPanel.tsx # Scores + did-well / missed / tip / tags
│ ├─ OutcomePanel.tsx # Outcome %, verdict, insight line, ideal answer
│ └─ ReflectionPrompt.tsx # 3-field structured journal entry
├─ data/
│ └─ scenarios.ts # 8 hardcoded scenarios (growth vs value, good vs bad)
├─ lib/
│ ├─ types.ts # Shared TS types
│ ├─ storage.ts # localStorage wrapper (single-key array)
│ ├─ scoring.ts # Verdict + progression math
│ └─ evaluationPrompt.ts # THE CORE PROMPT (see below)
├─ .env.local.example
├─ package.json
├─ tsconfig.json
├─ tailwind.config.ts
├─ postcss.config.js
└─ next.config.js
Setup
1. Install
cd "Investing App"
npm install
2. Configure the AI provider
Copy .env.local.example to .env.local and fill in ONE provider:
cp .env.local.example .env.local
Anthropic (default):
ANTHROPIC_API_KEY=sk-ant-...
AI_PROVIDER=anthropic
AI_MODEL=claude-sonnet-4-5
OpenAI:
OPENAI_API_KEY=sk-...
AI_PROVIDER=openai
AI_MODEL=gpt-4o-mini
No API key? The app still works — /api/evaluate falls back to a deterministic local rubric so the full flow remains usable offline. Feedback quality is worse; that's the point of using a real model.
3. Run
npm run dev
# http://localhost:3000
4. Build for production
npm run build
npm start
Environment variables
| Variable | Required | Description |
|---|
AI_PROVIDER | No | anthropic (default) or openai. |
AI_MODEL | No | Model name. Defaults: claude-sonnet-4-5 / gpt-4o-mini. |
ANTHROPIC_API_KEY | If anthropic | Anthropic API key. |
OPENAI_API_KEY | If openai | OpenAI API key. |
Nothing is ever sent to any third party besides the chosen AI provider. API keys stay server-side in app/api/evaluate/route.ts; do not expose them as NEXT_PUBLIC_* variables. All decisions, reflections, and progress live in the browser's localStorage under the key ait.decisions.v1. localStorage is readable by any script running on this origin, so XSS prevention remains critical and this data should not be treated as secret storage.
The AI evaluation prompt
The prompt (in lib/evaluationPrompt.ts) is the most load-bearing part of the product. It:
- Frames the model as a senior buy-side equity analyst mentoring a junior — direct, slightly critical, not a cheerleader.
- Grades along four explicit axes: use of valuation, growth vs profitability, risk awareness, logical reasoning.
- Explicitly penalizes vague answers, ignored metrics, and unjustified overconfidence.
- Forces a strict JSON output shape so the client can reliably render scores, bullets, tags, and an ideal-answer comparison.
- Requires every piece of feedback to reference something concrete from the scenario (a number, a headline, the signal).
If you want to tune the product's feel, tune this file.
Design principles honoured
- No auth, no real money, no DB. Pure client-side storage + one API route.
- Flow over features. The user sees exactly one thing at a time: scenario → decision → feedback → outcome → reflection.
- Feedback is specific. Generic praise is explicitly discouraged in the system prompt.
- Progression is earned. Skill score is a rolling average of decision-quality scores; accuracy ignores neutral outcomes; calibration measures whether confidence is backed by evidence.
Non-goals (deliberately out of scope)
- Real brokerage integration or live prices.
- Portfolio accounting, P&L, or Sharpe-ratio theatre.
- Free-text rich-editor journaling.
- Educational courseware.
Keep it focused. Reps + feedback is the product.
