skills

Scaffold and maintain a self-hosted Astro Starlight documentation site in the current repository, and wire the repo's agent file (CLAUDE.md / AGENTS.md / etc.) to keep docs fresh alongside commits and releases. Use when the user wants to "add docs", "set up a docs site", "scaffold Starlight", "create a documentation site", "add llms.txt", make docs AI-readable, or establish a docs-as-code workflow. Inspects the repo first (package manager, monorepo layout, existing docs, CI, agent files), decides non-bloating placement, applies design tokens, sets up easy LAN/self-host serving, and adds a docs freshness directive to the agent file. Builds the docs *shell* (site, tokens, llms.txt, CI, freshness wiring) with stub pages; pair it with the `harvest-docs` skill to populate the pages with comprehensive, code-accurate content.

Give a long-running, self-hosted tool its own lifecycle CLI — a single `<tool>` binary with install / upgrade / doctor / status / logs / restart / uninstall (and optional prune) subcommands, a managed service unit (systemd user unit on Linux, launchd plist on macOS), channel-based self-upgrade (stable = highest release tag, nightly/dev = branch tip), and a `/health`-probe that verifies the running version before an upgrade is called done. Use when the operator says "give this tool an install/upgrade CLI", "add self-upgrade", "channel-based upgrade command", "make it self-managing", "a factory-style CLI for this service", "born with the maintenance surface", or is about to hand-port a service-management CLI from another project. It implants the operational lifecycle surface; pair it with a release/versioning flow (tags + changelog) for the other half.

Operating mode for long-running autonomous sessions where the user is unavailable (overnight, traveling, AFK for hours) and cannot answer clarifying questions. Shifts Claude from "ask when unsure" to "decide, document, and keep moving" while preserving hard guardrails on irrecoverable actions. Use this skill ONLY when the user explicitly invokes it — phrases like "run overnight", "I'm going to bed, keep working on X", "autonomous run", "no need to check in until morning", "/overnight", or any direct request to operate without check-ins. Do NOT trigger on normal long tasks where the user is still at the keyboard. The user has explicitly accepted the cost of any recoverable mistakes; in exchange, expect Claude to ground decisions rigorously, log them auditably, and produce a morning report so the work is reviewable.

Cut a release for a Bun/Vite/tRPC-style project — propose a semver bump from conventional commits, write a Keep-a-Changelog entry, bump version(s) in lockstep, annotated-tag, and hand the push to the operator. Also the canonical reference for wiring the runtime "what's new" loop (changelog parser → API → version-gated modal) and a channel-based upgrade CLI into a new project. Use when asked to "cut a release", "tag a version", "bump the version", "update the changelog", "do a release", or to add release-notes / what's-new / self-upgrade machinery to an app.

Run a full strategic leverage audit on any repository and produce an executable improvement plan across four lenses — technical innovation, defensibility/moat, profitability (or open-source health metrics), and product/adoption metrics. Use this skill whenever the user points Claude at a repo and asks to "improve", "audit", "grow", "harden the moat of", "make more profitable", "make more adoptable", "find leverage in", or "level up" a project — or asks what to build next, how to position a project, how to get more stars/users/revenue, or whether a project is worth continuing. Also use it for pre-open-sourcing readiness checks and build-in-public preparation. Designed for frontier-model (Fable 5 class) deep-context analysis; do not answer these questions ad hoc when this skill is available.