shimo4228/claude-code-learned-skills Skills

Framework adoption decision matrix: custom vs large frameworks in the Claude Code era. Use when evaluating whether to adopt a large framework or build custom with AI.

Use when replacing a core algorithm (encryption, hashing, ML model) that affects persisted user data and needs rollback safety.

Use when extending a frozen dataclass or Pydantic pipeline with new fields without breaking existing consumers.

CJK/Latin weighted token estimation for multilingual LLM pipelines. Use when processing Japanese/Chinese/Korean text with fixed chars-per-token constants.

Use when adding MCP servers from within a Claude Code session where CLI is unavailable. jq-based ~/.claude.json editing workaround.

Use when writing large files (>1000 lines), editing after many operations, or configuring hooks in Claude Code.

Use when caching expensive file processing results. SHA-256 content-hash keying with frozen CacheEntry and service layer wrapper.

Use when building an LLM-powered app that needs cost control via model routing, budget tracking, retry, and prompt caching.

Use when building automated research systems. Official Deep Research APIs (OpenAI, Gemini, Perplexity) over browser automation.

Use when designing Swift model structs that need predictable, thread-safe state updates without mutation.

Use when validating auto-generated JSON data files too large for manual review. Layered pytest validation pattern.

Use when evaluating LLM-generated structured output against expected results using keyword matching and F1 metrics.

Use when processing documents over 50K characters through LLM APIs with section splitting and batch cost reduction.

Use when mock assertions fail due to public functions forwarding internal params like url or timeout to a shared helper.