// the find
badlogic/claude-commands
Global Claude Code commands and workflows
A pair of Markdown-based workflow prompts that structure human-agent collaboration for feature development: one using git worktrees for parallel isolation, one staying on the current branch. Aimed at engineers who want to keep their hands on the wheel while offloading the mechanical parts to Claude Code or similar agents.
The state machine approach (INIT → SELECT → REFINE → IMPLEMENT → COMMIT) gives agents a clear contract about what they're allowed to do at each phase, which reduces scope creep better than a freeform prompt. Zero dependencies — it's just Markdown, git, and shell, so it ports to any agent without adapter code. The explicit separation of the research artifact (analysis.md) from the implementation plan (task.md) from the todo list is a sensible data model for resumable work. The README is unusually honest about failure modes, including LLM unreliability and the race condition when multiple humans share todos/todos.md.
At ~150 LOC of Markdown, the workflows live or die on whether the LLM follows state transitions faithfully — there's no enforcement mechanism, so a confused model mid-task can silently skip phases. The '--dangerously-skip-permissions' flag in the recommended invocation is a significant surface area to hand to a model executing shell commands; it's acknowledged but glossed over. There's no mechanism to validate that the generated task.md plan actually maps to the original todo intent before IMPLEMENT begins — a confident but wrong refinement phase just becomes a confidently wrong implementation. The project is essentially a single person's workflow preferences published as Markdown; the 31 forks suggest some adoption but there's no community contrib process, issue triage, or versioning, so it may quietly bitrot as Claude Code's own slash-command features evolve.