learning-checkpointlisted

# Learning Checkpoints — Common Applies to every project type. Triggers happen every task, independent of doc/validator timing (see Sprint Documentation Sync) — this is about live discussion in the session, not a file that gets written and synced later. **Claude Code note:** Checkpoint A's "ask, don't guess" step and Checkpoint B's item 1 are conversational — they depend on the agent choosing to follow this file. If `adapters/claude/pretooluse_scope_guard.py` is wired into `.claude/settings.json`'s `PreToolUse` hook, the same "ask before implementing" rule is also enforced mechanically: it blocks `Edit`/`Write`/`MultiEdit`/`NotebookEdit` on source files until `current-state.md` has a scoped `Current Task` and a filled `Clarifying Questions Asked` field. That hook is a backstop, not a replacement for actually running the checkpoints below — it can't tell whether real questions were asked, only whether the field was filled in. --- ## Checkpoint 0 — Unfamiliar Technology (only when it applies) Trigger: the current task touches a technology, library, or pattern you have never used before — not just "existing code in a language I know," but genuinely new ground. When unsure whether this counts, run it anyway — a false-positive Checkpoint 0 costs one short grounding pass; skipping a real gap means building on an assumption that was never checked. Run this *before* Checkpoint A or B, as a grounding pass: 1. "用我熟悉的東西打比方，這個技術是在解決什麼問題？" — force an analogy to something already kn