Claude Post but Useful (playtechnique.io)

A developer turned a bunch of “vibe coding” blog-post claims into a repeatable Claude-driven dev workflow by teaching Claude Code explicit, natural‑language “slash commands” that act like small agents. They created .claude/commands/todo and /emergency handlers that read a README → generate a todo.md, pick the next unfinished item, implement it, run tests, and use precommit/commit agents to commit changes. The emergency command enforces a tests‑first rule (write a failing test that reproduces the bug before fixing it), and the author added a notifier hook in ~/.claude/settings.json to surface when Claude is waiting for permissions or user interaction. Technically this shows two practical patterns for LLM-assisted development: (1) “programming” the model via natural‑language command recipes (persona + step list) rather than ad‑hoc prompts, and (2) constraining tool use and side effects with explicit CLI behaviors (e.g., use git diff --no-ext-diff to stream diffs). Important operational details include granting Bash tool permissions carefully, the quirks of the settings wildcard, and hooking Notification events to desktop alerts (terminal‑notifier). The result is a fast prototyping loop that materially speeds feature work but still produces buggy output and requires human review and security caution — useful as an assistant and automation layer, not an autonomous replacement.