← ClaudeAtlas

runlisted

Drive a GitHub issue — bare or tracked on a Project board — from triage to a merge-ready PR through a gated pipeline (design hardening, tests green, code-review clean), scaling the machinery to the task's tier and asking at most one batched question. Auto-links the issue to close on merge, advances the board card, then merges and cleans up once you approve the PR in-session. Triggers: "take task N", "work on issue #N", "build/fix X" when no issue exists yet, and — for the merge gate later — "merge it", "approve the PR", "ship it", "lgtm merge".
DmitriyYukhanov/claude-plugins · ★ 7 · Code & Development · score 76
Install: claude install-skill DmitriyYukhanov/claude-plugins
# issue-to-pr — issue → merge-ready PR pipeline Ten steps from an issue to a merged PR: the gates block, the depth between them scales to the tier. Scripts own the mechanics (`S/` = `${CLAUDE_PLUGIN_ROOT}/scripts/`, `R/` = `references/`), you own judgment. One todo per step. Any non-zero script exit → `R/contracts.md` before you act: some are stops, some are fallbacks. ## Hard rules (never violate) - **Merge is gated on explicit in-session approval**, runs ONLY in the main session, via `S/worktree.sh merge` — never a bare `gh pr merge`, which no hook denies, and never on the turn the PR opens. - **Ask contract:** three moments, `--grill` reshapes the first (`R/judgment.md`) — (1) Step 3: ONE batched `AskUserQuestion` if the ledger has open items, or the grill in its place, (2) the merge gate, (3) a hard stop. Decide everything else yourself and log it, and never ask what a script or the code can answer. - Stage with **explicit paths** (`git add path1 path2`); never `git add -A`/`.`, which sweeps in whatever the project keeps untracked. Write multi-line code with the Write tool — a Bash heredoc breaks on an apostrophe inside the body. - **Evidence before assertion:** no "green/passing" without the command output; look up any claim about the world outside this repo before building on it, ledgered either way (`R/judgment.md`). - **Humanizer** runs on 100% of human-facing text (report, PR body, UI strings > 1–2 words) regardless of tier — not code, logs or co