pr-shepherd
xdanger/dotfiles
End-to-end driver that takes a ready-for-review pull request all the way to a clean merge. Use whenever the user wants Claude to shepherd, babysit, watch, or drive a PR to green and merge it — e.g. "take PR #123 through CI and merge it", "babysit this PR until everything is green", "land this PR and clean up". It loops over the PR's checks and reviews: fixes real CI failures (re-running only genuine flakes), replies to every review comment, resolves each conversation, and re-polls until a strict gate passes — all checks actually completed and passing, every thread resolved or outdated, mergeable with no required approval outstanding — not merely "looks fixed". Then it merges with the repo's preferred method, deletes the remote branch unless it is long-lived, removes the local branch and git worktree, and fast-forwards local main/master. Trigger even when the user only describes the outcome ("get this merged once CI is happy") without naming the steps.
deep-dive
xdanger/dotfiles
Structured pedagogical deep-dive for new substantive topics across STEM, technical/engineering practice, and humanities/business/history. Use this skill whenever the user asks to learn, understand, deep-dive, explore, or be taught a new concept they don't already have a working model of — including phrasings like "what is X", "how does Y work", "why does Z happen", "explain W", "带我入门 X", "讲讲 X", "教我 X", "我想搞懂 X", even when they don't use the word "teach" or "learn". Trigger especially when the topic is a named concept, framework, theorem, mechanism, technology, historical event, or intellectual movement. Do NOT trigger for quick factual lookups ("when did X happen"), tool/syntax questions ("how do I curl with auth"), conversational chitchat, or mid-flow technical discussion where the user is already engaged on the topic.
audience-aware-comms
xdanger/dotfiles
Apply whenever you are about to produce written output that an INTERPRETING READER — a human, or another AI agent that will act on your words — will read and act on: emails, IM/Slack, PR or issue descriptions, review comments, design or decision docs, customer-facing copy, AND natural-language task specs, prompts, or agent mandates whose executor is a person or an LLM. Use it even when the user did not ask for "communication" and the message is only a sub-step of a larger task (e.g. the PR body after a feature, or a prompt you are writing for a downstream agent) — those incidental cases are where audience-blind writing slips through, so reach for this proactively. It makes you model the reader's actual capability and what they will infer or assume, then emit only the calibrated artifact. Do NOT use for output consumed literally by a deterministic interpreter — executed code, configs, schemas, queries, regexes, test fixtures — where correctness is binary and there is no mind to model.
