← ClaudeAtlas

impl-pipelinelisted

Drive Issue implementation with support agents — plan (read-only), split mechanical vs. interactive at a human gate, implement, test from spec, review with review-team-run, triage, fix, and stop at a final human gate before /git-pr-create.
tatsushige-i/shared-claude-code · ★ 0 · Data & Documents · score 63
Install: claude install-skill tatsushige-i/shared-claude-code
# Implementation Pipeline Skill Orchestrate the post-planning implementation of a single GitHub Issue using a small team of support agents, while keeping every judgment call — the mechanical/interactive split, the review triage, and the go-ahead to commit — with the human (main session). This skill runs **after `/git-issue-start`** (branch created, Plan Mode plan approved) and **hands off to `/git-pr-create`** at the end. It reuses the existing `/review-team-run` skill for the review phase. ## Design principles (do not violate) - **The planner does not decide.** `impl-planner` presents the plan and the decision points; the human makes the mechanical-vs.-interactive call and every product/UX decision. - **coder implements a confirmed plan only.** Autonomous implementation (`impl-coder`) runs only for the mechanical parts, after the human gate. - **tester writes from the spec.** `impl-tester` derives tests from acceptance criteria / `docs/spec`, never from the just-written implementation code. - **Review findings are not fixed unverified.** Triage `review-team-run` output by severity and fix only CONFIRMED / worthwhile findings. - **No orchestrator agents, no agent nesting.** Orchestration is this skill's (main session's) job — do **not** create `impl-orchestrator` or `review-orchestrator` agents, and never have one agent launch another. Only skill→agent calls (this skill launches `impl-planner`/`impl-coder`/`impl-tester`; `/review-team-run` launches the review team). - **C