engineering-workflows

Six skills and one agent, covering the week between "something is wrong" and "it shipped".

root-cause runs a debugging session as a loop rather than a guess: symptoms first, then falsifiable hypotheses tested against evidence. The heavy reading is delegated to the debugger agent, which is read-only — it never edits and never runs a mutating command, so the diagnosis comes back to the main thread where the full test output is visible and the fix can actually be verified.

design-intake classifies the work before anyone writes code — a spike, a bounded change, or something architectural — and the ceremony scales with that answer while the approval gate before implementation does not. writing-plans turns an approved design into a plan someone else can execute: a pointer back to the spec, the constraints reproduced verbatim so a per-task dispatch carries them, and per-task file and interface blocks.

release-train runs a large change as one leader supervising several workers, each in its own git worktree, converging on a shared release branch. It reads the file blocks from a plan to prove the bundles are genuinely disjoint before any worker starts, because two workers editing one file is the failure the whole structure exists to prevent.

repo-walkthrough orients you in an unfamiliar codebase to the point of being able to contribute safely. handoff writes the note a fresh session can resume from — including what was already tried and ruled out, which is the part that stops the next session repeating the last one.

Everything here needs a real checkout and a shell, and the debugger agent is a subagent, so this plugin is Claude Code only.

If you want the full review pipeline — deterministic detectors, judgement agents and a validator that owns the verdict — that is code-review-core. If you want the Claude Code authoring rules those agents are written against, that is claude-craft.