bgutschke
- 9 skills
- 0 followers
- 17 hours ago last updated
- ▌ To Pr · bgutschkeInvoked by a human typing /to-pr, or by an in-session subagent whose own given task already implies landing its work as a PR. Open a new PR from the current branch (draft by default for a human, always draft when subagent-invoked; title and body derived from its commits and diff), or fill in an already-open PR's description using the target repo's own .github/PULL_REQUEST_TEMPLATE.md — falling back to a built-in What changed/Why/Testing structure. --ready/--draft toggle draft state; --base sets or retargets the base branch — see "Agent-invocation guardrails" for how a subagent's use of these two paths narrows.
- ▌
- ▌ Repo Catchup · bgutschke bundleReads the commits in the current repository for a date range. Replies with a one-line executive summary and a Date, Owner, Ref, and Description table. Drops bot-authored and merge commits. Groups rows by team, with a CODEOWNERS file present, otherwise by individual contributor. --from and --to each accept an absolute YYYY-MM-DD date, a relative phrase such as "yesterday" or "last week," or a git tag or branch name. When neither is given, defaults to the last seven days. Runs git log inside a delegated subagent, so a large range never floods the conversation. On a GitHub remote with gh authenticated, each row links to its merged pull request, or otherwise to its bare commit. Elsewhere it falls back to a plain hash. When the user types /repo-catchup, use this skill. When the user asks, in plain language, to catch up on this repository, use it too. Example requests: "catch me up on this repo," "what changed here recently," "summarize commits since last week."
- ▌ Audit Rules · bgutschkeReads every active rule file and installed skill/agent description, and reports contradictions and unresolved overlaps between them, proposing edits only to files the user can edit directly. Use when the user asks to audit, review, or check their rules, CLAUDE.md files, or installed skills/agents for contradictions, conflicts, or overlap, or types /audit-rules.
- ▌ Curate Memory · bgutschke bundleRuns one curation pass over the current project's memory store — reads the newest session transcripts, filters them to real user and assistant prose, and writes a complete candidate store beside the input plus one report listing every merge, drop, rewrite, and addition with its justification and a citation to the session that supports it. Every retained memory naming a concrete file, command, or flag is also checked against the working tree and reported verified or unverifiable — a failed check only annotates the report, it never drops the memory. The input store is never opened for writing, and the pass adopts nothing — it prints the command to adopt the candidate store and the command to discard it. Use when the user types /curate-memory, or asks to curate, consolidate, clean up, de-duplicate, or prune their memory store or memory files.
- ▌ Renovate Triage · bgutschke bundleReads each open Renovate dependency-bump PR's changelog, release notes, and CI status, computes a Risk verdict — safe, needs-review, or blocked — from a fixed hard-stop rule list, and posts (or updates) one PR comment stating the verdict and reason; blocked PRs also get an Agent brief naming call sites and changelog sections for follow-up. Resolves each changed file's datasource (npm, Docker, PyPI, Ansible Galaxy) from the repo's own renovate.json rather than guessing from filenames. For a minor or major bump it also reports relevant Opportunities in a separate, verdict-independent section (skip with `--no-opportunities`); for every bump it also scans gathered changelogs for a Security advisory — a CVE, GHSA ID, or urgency language — reported the same way, never changing the verdict. Use when the user types /renovate-triage with no argument to scan every open Renovate PR in the current repo, with a PR number or URL to check one, or asks to triage, review, or assess Renovate PR risk.
- ▌ Skill Writing Standards · bgutschkeInvoked via /skill-writing-standards to research external, authoritative sources on Agent Skills authorship (Anthropic's own docs plus other maintained skill collections) and reconcile the findings into docs/skill-writing-best-practices.md as proposed, human-reviewed edits.
- ▌ Refactor Rule Tree · bgutschke bundleRuns one placement-and-pointer pass over a rule tree — the personal global rule file by default, or a given root path, plus every file reached by import or mention edge. A rule in an auto-loaded file gets a verdict: stay, move to a topic file, become a model-invoked skill or hook, or delete. A file reached without being auto-loaded is left untouched — only its pointers get checked. Every pointer gets live, dead, or unverifiable; an uncited target is reported unrouted, separately. Before a move, finds every in-scope file that cites the rule's location and either updates those citations in the same change or blocks the move. Presents one complete plan, checked for completeness, before writing anything. Never compares two files' guidance against each other. Use when asked to refactor, restructure, reorganize, or clean up a CLAUDE.md or rule file, or when /refactor-rule-tree is typed.
- ▌ Draft Commit Message · bgutschke bundleDrafts a Conventional Commits message for the currently staged changes, matched to the target repo's own discovered commit-message convention (commitlint config, written doc, or git-log pattern) where one exists, falling back to a generic standards-based convention otherwise. Reads the diff and branch inside a Haiku subagent so the raw diff never enters the main session. Never runs git add or git commit, even when the request is to commit changes outright — it only returns message text for review. Use when the user types /draft-commit-message, explicitly asks to draft, write, or generate a commit message, or asks to commit changes outright ("commit this", "commit these changes") — in the last case this skill supplies just the message-drafting step, leaving staging and committing to the harness's own default flow or a separate explicit request.