SKILL

Turn a completed non-trivial engineering session into a paired durable artifact: (1) a reusable skill under Innei's personal SKILL repo, and (2) a published blog post that distills the lessons, linked to the skill via the attached skill card (meta.skillIds — never mentioned in the body). Triggers on "把这个过程写成 skill 再写一篇 blog"、"沉淀一下这次的折腾"、 "productize this session"、"publish this as a skill and a writeup".

Use when extracting an in-app Electron native-module integration (N-API addon + build scripts + packaging config + release CI) into a standalone, source-distributed npm library — or when authoring such a library from scratch. Covers the ABI/install-script traps, node-gyp rebuild semantics, macOS rpath depth, electron-builder packaging excludes, composite GitHub Action hygiene, and the dogfood migration back into the source app.

Use when rolling out a rewritten site (new framework, new Vercel project) behind an existing production domain gradually — cookie-sticky percentage split between two origins via a Cloudflare Worker on the zone route, with hashed build assets offloaded to R2 so asset traffic never invokes the worker. Covers why middleware-based canaries break CDN caching, the same-zone-subrequest HTTP trap, cross-bucket cache poisoning, and an R2 layout that survives many concurrent versions.

Use when the user asks for a work summary, weekly report, 工作总结，周报，working summary, or sprint/period recap. Aggregates GitHub PR/commit/issue activity from configured repos, optionally reads Linear cycle issues via MCP, honors Chinese public holidays, and produces a markdown report. Default range is the previous Mon-Sun week.

Use when migrating a multi-locale Next.js RSC / App-Router site to React Router v8 so SSR HTML is CDN-cacheable. RSC payload is uncacheable at the edge and locale × page count multiplies the bill. Covers ordered cuts: strip UA/geo/cookie SSR variance, split `.data` loaders from client fetches, restage theme SSR (next-themes + head script + /theme-vars.css), split middleware to Vercel Routing Middleware + root loader mirror, prerender static, land 3-tier Cache-Control so Cloudflare owns cache.

Use when generating a LINE/WeChat-style chibi sticker sheet (4x8 grid, 32 expressions) from an anime character reference image via Gemini, including transparent PNG output and individual cell slicing.

Use when deploying VLESS+Reality+Vision on AWS Lightsail and bridging it to Surge (or similar TUN clients) with an Alpine LXC SOCKS5 bridge on the LAN plus a brew sing-box on the laptop for off-LAN use. Triggers on "set up vless reality on aws", "tokyo lightsail proxy", "surge doesn't support vless", "vless when away from home lan", "alpine lxc client for vless".

Access the remote mx-space PostgreSQL through ssh to the swarm host, then docker exec into the Postgres container and run psql inside the container. Use for inspecting tables, sampling rows, validating topic assignments, and performing guarded updates in the specific mx-space deployment pattern where direct host-level Postgres access is unreliable.

Audit remote mx-space translation data through ssh to the swarm host, then docker exec into the Postgres container and run psql inside the container. Use for checking translation_entries coverage, ai_translations gaps, strict computeContentHash mismatches, and runtime freshness semantics in deployments where direct Postgres access is unreliable.

Use when you need to canary-deploy two HTTP backends on the same Dokploy-managed domain with a percentage split (e.g. migrating an SPA from one framework to another). Covers the critical SPA-asset trap that breaks naive sticky-cookie splits, a path-aware Traefik config that solves it, and a four-phase workflow (backup → dry-run on a canary host → roll to prod → quick rollback).

Use when user provides exported chat history (WeChat / Telegram / iMessage / QQ exports as .md / .txt / .json) and asks "我和 X 聊了啥"、"看完整个 md"、"细说 XX"、"是不是有 Y"、"what did I and X talk about", wants topical breakdown, requests detail on specific themes (relationships / work / health / events), or seeks honest interpretation of conversation dynamics. Triggers on large chat dumps (>1000 lines) where direct full read is impractical.

Use when a runner, CI step, container, or orchestrator executes a task whose stdout/stderr you cannot retrieve afterwards — no log API, no shell access, output stream not captured, or logs evicted before you can read them. The pattern: have the task itself write its captured output into a data store you CAN read (SQL row, Redis key, mounted file, public webhook).

Use when a CI release/build pipeline runs a smoke test that boots the project's bundled binary against an external dependency stack (PostgreSQL, MySQL, Redis, MongoDB, S3, etc.) and the test fails with connection refused, missing-required-env, or migration errors. The fix is to declare the dep as a CI service container AND inject the matching env vars on the smoke step. Common after a stack migration leaves the release workflow out of sync with the new ci.yml/code reality.

Use when operating a Dokploy-managed deployment via REST API from the shell — creating/updating/deploying postgres/redis/compose/application services, switching an application's source between docker registry and git build, or scripting redeploys. Covers auth, the working endpoint catalog, and a critical gotcha about deployment status semantics.

Use when running an ephemeral one-shot task (data migration, batch job, schema setup, ad-hoc psql script) that must reach other services in a Dokploy project's internal network without exposing those services to the public internet. The task should join the project's docker overlay network, run to completion, and exit. Pairs with dokploy-api-cli.

Produce a self-contained handoff prompt for another agent (Codex, a fresh Claude session, a teammate) when the user wants to delegate continued work. Triggers on: "写一个 prompt 给 codex"、"交接一下"、"让 xxx 跟进"、"summary 一下再写个 prompt"、"handoff to another agent"、"write a prompt so X can continue", or any request to capture the current session state for continuation elsewhere.

Use when delivering analysis, assessment, or diagnosis to the person whose work/behavior/decisions are being analyzed, and they push back emotionally, contradict conclusions, or express discomfort — prevents flipping the conclusion to match their mood while staying open to new factual evidence. Covers code review, architecture review, performance diagnosis, root-cause analysis, decision post-mortems, behavior/relationship analysis — any task where the reader is the subject.

Use when generating an ACG (anime/manga) character settei sheet (キャラ設定資料) — multi-view full-body lineup, expression list, detail callouts, color palette — from a template settei image plus a single character reference image. Built on top of gemini-image-generation; this skill encodes the layout-locking + identity-locking prompt pattern that reliably produces settei-style output.

Use when a task requires Gemini text-to-image or image-to-image generation, including style transfer, character consistency, reference-image workflows, and watermark removal from Gemini-sourced images.

Use when a web project needs Google AI Studio or Gemini-generated favicon and Open Graph images, exported icon variants, and matching SEO metadata wiring.

Use when uploading one or many files to a Cloudflare R2 bucket and returning public URLs via a bound custom domain. Covers wrangler CLI and the REST API (Cloudflare API token) paths, batched uploads, MIME handling, correct account selection for multi-account setups, and post-upload public-URL verification. Triggers on tasks like "upload these images to R2", "push assets to the uploads bucket", "put files under <prefix>/ on R2".

Generate a DESIGN.md file for any brand or website by analyzing its visual design system. Use when the user asks to "generate a DESIGN.md for [brand/URL]", "create a design system doc for [site]", "extract the design tokens from [URL]", or "make a DESIGN.md like [brand]". Triggers on any request to document a website's design language in DESIGN.md format.