No-code investor builds 2 Japan stock tools with Claude Code

The two tools are deliberately opposite in character. 投資可否チェッカー starts from financial data and earnings disclosures, computing a 100-point score split across growth, profitability, financial health, cheapness and stability, and adds price action and human-reviewed 四季報 articles as separate layers. 急騰株チェッカー instead watches 15 themes such as cybersecurity, space and defense every day, scoring signals like US peers rising first, a sudden jump in related TDnet disclosures, or volume-backed surges spreading among domestic names. Its theme tickers and keywords live in JSON config files, so themes can be added without touching code.

The design principle behind both is that the AI is not allowed to judge. Numbers, technicals, stop-loss levels and the final verdict all come from fixed rules, and the reasons are framed as two: traceability, since every number, condition and reason is stored so a bad call can be traced back, and safety, since missing data is never filled in by guesswork and suppresses strong buy signals. AI drafts are kept separate from approved content and never overwrite them until a human signs off.

The development setup mirrors that discipline. The investor only sets policy and investment rules and gives approvals; Claude Code acts as PM, splitting, assigning and checking work; two sub-agents handle scoped implementation and money-, share-count- or stop-loss-related changes. Roughly 45 spec documents in docs/ are treated as the source of truth, undecided rules sit in a dedicated document the AI may not resolve on its own, and CLAUDE.md limits what the PM must ask about to three cases.