Shepherds a PR through the repetitive loop: run qa-swarm review and the simplify pass in convergence rounds, triage the review conversation (review-triage), keep the branch current and report CI (ci-shepherd), apply the `stamphog` label on each new commit, and report stamphog's verdict. Use when the user says "/pr-shepherd", "shepherd this PR", "babysit this PR", or wants the whole review loop driven automatically. Accepts an optional PR number or URL as argument.
Triages the review conversation on a PR: runs qa-swarm when there are substantive changes, classifies every unresolved review thread (qa-swarm, AI/bot, and human) as actionable / nit / ambiguous, applies the clear single-file fixes, resolves nits silently (reasons go in the report), runs ambiguous AI/bot threads through paul-pair's autonomy ladder before deferring so only genuine stop-and-ask cases reach a human, and defers any thread with human participation. It never posts thread replies — responses to humans always come from the PR author. Use when the user says "/review-triage", "triage the review comments", "deal with the bot comments", or "address the review feedback". Accepts an optional PR number or URL as argument. Does not touch the branch base or the stamphog label.
Orchestrates four review skills (qa-team, paul-reviewer, xp-reviewer, security-audit) into a single comprehensive PR review with inline GitHub comments. Use when the user asks for "qa-swarm", "swarm review", or wants a full multi-perspective review posted to their PR. Accepts an optional PR number or base branch as argument.
Keeps a PR's branch current with its base (fast-forward trunk + restack via the Graphite MCP, with a trivial-vs-needs-decision conflict-resolution sub-workflow) and reports CI state without ever gating on it. Use when the user says "/ci-shepherd", "restack this PR", "is my branch behind?", "keep this branch current", or "what's CI saying?". Accepts an optional PR number or URL as argument. Reads no review or stamphog state.
Pair-programming with Paul D'Ambra's engineering judgement — useful to any engineer who wants a take-ownership default instead of asking permission for every improvement. Run it at two moments: BEFORE asking your pair a clarifying or permission question, and BEFORE declaring a task finished. It decides whether you can act autonomously, should recommend-and-ask, or must stop-and-ask, and forces the "am I done, or should I make this better?" refactor checkpoint that is the third beat of every red/green/refactor cycle. Use it whenever you catch yourself about to type a question to your pair, about to say "done"/"this is ready"/"finished", or about to pass up a simplification because nobody asked for it. See references/worked-examples.md for concrete preferences and the principle each one serves.
Pick up a stale or abandoned PR that someone (often you) started and stopped, bring it back to a shippable state, and hand it off to pr-shepherd for the active review loop. Use when the user says "take over this PR", "takeover and complete this PR", "continue <PR URL>", or pastes a PR URL and asks to finish it. Accepts a PR number or URL as argument. Distinct from pr-shepherd: the PR is dormant, not in an active review cycle.
Hunt browser memory leaks using all available signals — off-heap renderer process growth (phys_footprint/RSS), detached DOM element counts, heap snapshot diffs, network activity, CDP performance metrics, and memory-infra allocator traces. Covers the multi-tab Playwright harness, production cmd-click testing, memlab MCP heap analysis, in-browser __leakHunter helpers, CDP forced-GC, Apple footprint tool, and memory-infra traces. Use when the user says "browser tabs use gigabytes", "renderer RSS growing", "phys_footprint climbing", "memory leak", "tabs slow/crash", "hunt detached elements", "investigate react memory leaks", "find detached nodes", or asks to investigate PostHog's memory problem.
Keep Claude memory backed up by storing the source of truth in ~/dotfiles/ai/claude-memory/ and symlinking each project's memory dir from PostHog Code's app storage to the dotfiles copy. Use this skill BEFORE saving any memory file in a project where the memory dir is not yet a symlink — the auto-memory instructions in the system prompt will otherwise write into the local-only Code storage and the memory will be lost on disk failure. Also triggers on "back up claude memory", "set up memory on a new machine", "restore claude memory", or when manage-claude-memory or link-memory.sh is referenced directly.