skills

Operate Linux servers over SSH: Debian/Ubuntu systemd setup and maintenance, distro-aware health and intrusion audits, SSH, firewall and container exposure, swap, and proxy/VPN setup or tuning including Snell repair. Use for unclear or possibly compromised hosts as well as known hosts. Not for macOS network problems; Snell evidence-only audits belong to surge, and the REALITY+HY2 stack belongs to sing-box-reality-hy2.

Start or continue a practical Web full-stack app on a React + TanStack Router + Vite + Convex stack — a dashboard, SaaS, internal tool, or product prototype needing a coherent stack choice, scaffold command, auth/data boundary, and first usable feature. Not for isolated changes to an existing feature, and not for Next.js/SSR/SEO sites, React Native, or payment flows; a live-built Next.js Convex app belongs to convex:quickstart.

Read-only local macOS Things queries and summaries when the user asks about their Things tasks, Today list, Inbox, projects, tags, completed or logbook history, or task search. Read-only, never writes.

Deploy, repair, and validate a narrow sing-box stable stack: VLESS REALITY Vision on TCP/443 plus Hysteria2 on UDP/443. Use for Debian/Ubuntu systemd VPS setup, Linux system-level TUN, cross-platform client configs, migration off Clash/Mihomo, REALITY/HY2 testing, or a 科学上网/翻墙/梯子 request when the user is open to this REALITY+HY2 stack. Surge-related macOS network triage and Snell audits belong to the surge skill.

macOS network triage when Surge is installed, active, or plausibly in the network path: Surge CLI, profile syntax, Enhanced Mode / System Proxy / DNS failures, policy routing and smoke tests, Tailscale coexistence, and Snell VPS evidence audits with repair plans. Server-side execution belongs to linux-server; REALITY+HY2 deployment belongs to sing-box-reality-hy2.

Set up scheduled agent work where the agent itself runs each check and nothing is deployed: do something later, keep checking until something changes, watch for something to finish and report it, or follow up periodically. Use when the user asks to schedule, loop, keep watching, or be reminded later — decide the scheduling shape first, then create it with the harness's own scheduling mechanism. Not for deploying durable monitoring infrastructure — that is end-to-end-monitoring.

Use when the user wants to review or summarize Feishu/Lark meetings over a date range, such as 总结这两周的会 or a meeting weekly report: find minute_tokens, pull transcript files, inspect coverage and duplicate evidence, build Chinese prompts, and optionally fan summaries out to Amp. Do not use for a single minute_token or single-meeting lookup (that is the lark skill), or for turning chat history into a report corpus — that is lark-chat-corpus.

Use when a task needs AI-generated or AI-edited bitmap assets, visual QA of generated images, screenshot-backed image work, UI annotations, or the local image workbench CLI. Not for deterministic local edits such as crop, resize, or compress, and not for diagrams, charts, or images that must carry precise text or code — those belong in vector tools such as SVG/HTML, Typst, or Mermaid.

Author, test, and maintain ast-grep rules and project lint gates: rule YAML capabilities, test-first authoring, suppression semantics, exit-code gating, pattern pitfalls, and version pinning. Use when writing or debugging ast-grep rules, wiring sgconfig/scan into pre-commit or CI, banning or enforcing a code pattern repo-wide (custom lint rules, even when the user never says "ast-grep"), or doing structural search/replace beyond a one-line pattern. Not for plain text search (use rg), one-off obvious patterns the CLI one-liner already solves, or the programmatic APIs (napi/pyo3).

Build a rich, self-contained interactive HTML explainer for a code change, diff, branch, or PR: layered background, intuition with diagrams, a code walkthrough, and a self-check quiz. Use when the user wants a durable, shareable explanation page. Not for an in-chat explanation or summary, regardless of diff size, and not for slide decks.

Design, deploy, repair, or verify standing monitoring and alerting infrastructure that runs on its own, independent of any agent session, for a host, service, cron job, crawler, dependency, data pipeline, or AI workflow. Not for the agent itself periodically re-checking something and pinging you — that is scheduled-work.

Send, test, or integrate brrr.now push notifications when the user wants to be pinged, reminded, or woken outside chat. Not for implementing push notifications inside the user's own product.

Audit product ideas, feature requests, startup directions, MVPs, and demand claims for real user demand, adoption risk, evidence quality, delivery feasibility, and concrete validation steps, including designing non-leading user interviews. Delivers a go/no-go verdict, not a cited research report. Not for stress-testing how to execute a plan the user has already committed to.