Hayao0819

Drafts fact-based technical articles for hayao's outlets — Zenn (@hayao), Qiita (@Hayao0819), the personal blog (hayao0819.com) — matching his structural register without faking his personality. Use when writing, drafting, or rewriting a 技術記事 / blog post / Zenn / Qiita article in hayao's name about a technical topic. The LLM writes facts only: it does NOT manufacture his jokes, 毒舌, 自虐, casual throwaway closers, or first-person anecdotes — performing his personality is fabrication and he finds it offensive. Encodes the factual baseline of his writing (genre-driven 敬体/常体, direct openings, plain content-label headings, inline-code density, primary-source links) in a bundled reference.md, and composes with the natural-writing skill. Self-contained — does not depend on machine-local memory. Do NOT use for: non-hayao docs, pure code, commit messages, or chat replies.

Rewrites prose so it reads as if a human wrote it — stripping out LLM tells in both English and Japanese. Use whenever producing user-visible writing of any kind: documentation, README, markdown, blog posts, commit messages, PR descriptions, GitHub Issues, code-review comments, code comments and docstrings, replies in issue threads, professional or casual emails, Slack/Discord/chat messages, release notes, design docs, academic-style writeups and 論文. Use it even when the user did not ask for "natural" prose — any drafted text leaving Claude that a human will read should pass through this checklist first. Triggers include: any request to "write", "draft", "compose", "summarize", "explain", "describe", "review", "comment", "reply", or 「書いて」「文章作って」「まとめて」「メール作って」「返信して」「PR/Issueの説明書いて」. Do NOT use for: pure code generation with no prose, raw data transformations, or terse tool output.

Security audit of memory-safety vulnerabilities in C/C++ (userland/systems, not kernel), with confidence-gated reporting to keep false positives low. Taint-traces untrusted input (argv, env, files, network, stdin) to memory sinks (copy length, array index, allocation size, format string) and reports only when input is attacker-controlled AND reaches the sink AND bounds aren't enforced. Covers buffer overflow, use-after-free, double-free, OOB read, integer overflow into undersized alloc, format string, the dangerous-function table, hardening flags to verify, and the ASan/UBSan/MSan/fuzzing toolchain. Covers userland/systems C and the SGX enclave/host (ECALL/OCALL) pointer boundary. Use when reviewing C/C++ for memory-corruption bugs. Kernel code → auditing-kernel.

Security audit of Linux kernel code and out-of-tree modules/drivers, with confidence-gated reporting to keep false positives low. Traces attacker input from userspace entry points (syscalls, ioctl, sysfs/proc/debugfs writes, netlink, char-dev, mmap) — and, for confidential-computing guests, from the untrusted host — to memory/refcount/locking/infoleak sinks. Covers UAF via refcount imbalance, the copy_to/from_user boundary, double-fetch TOCTOU, missing capability checks, integer-overflow allocations, ioctl handler review, and the TDX/SEV-SNP/SGX threat model. Use when reviewing kernel C, a driver/ LKM, an ioctl handler, or enclave/attestation code for vulnerabilities.

Security audit of web applications — backend implementation and authentication/ authorization flows — with confidence-gated reporting to keep false positives low. Traces attacker-controlled request input (params, body, headers, cookies, path) to sinks, distinguishing server-controlled from attacker-controlled, and reports HIGH only. Leads with broken access control / IDOR (OWASP #1), then injection (and where GORM/sqlc/React are already safe), XSS, SSRF, CSRF, auth (password storage, sessions, reset), and tokens (JWT alg confusion, OAuth/OIDC PKCE). Framework-aware for Go/Gin/GORM/sqlc and Next.js Server Actions. Use when reviewing a web backend, API, route handler, Server Action, or auth flow.

Idiomatic, version-current Go review and authoring for an experienced engineer. Covers deliberate error wrapping (errors.Is/As/Join, AsType), structured concurrency (errgroup, context, the loopvar fix), modern stdlib that replaced old patterns (slices/maps/cmp, log/slog, min/max/clear, range-over-func iterators), table-driven and synctest testing, API design the linter can't check, and Gin/GORM/cobra backend patterns. Use when writing, reviewing, or refactoring Go — .go files, go.mod, goroutines, errors.Is/As, a Go backend or CLI — or before committing Go. Flags outdated pre-1.22 idioms (loopvar v:=v shadows, tools.go, hand-rolled helpers).

Idiomatic, version-current Kotlin/Android review and authoring for an experienced engineer, Compose-first on Kotlin 2.x (K2). Covers coroutines/Flow pitfalls (GlobalScope, hardcoded dispatchers, exposed MutableStateFlow, collectAsStateWithLifecycle, stateIn), Compose idioms the linter misses (state hoisting, strong skipping, effect keys, derivedStateOf, modifier order), Kotlin language idioms (!! and lateinit, sealed + exhaustive when, value classes, kotlinx.serialization over Gson), and the official UI/domain/data architecture with Hilt. Use when writing or reviewing Kotlin or Android — .kt files, Composables, ViewModels, Gradle/version catalogs, coroutines — or Jetpack Compose UI. Pairs with the android-control skill for on-device work.

Idiomatic, version-current Next.js App Router review and authoring for an experienced engineer on Next 16 / React 19. Covers the 'use client' module-graph boundary and serializable props, Server Actions treated as public unauthenticated endpoints (validate + authN + authZ inside every one), the caching model that flipped across Next 14→15→16 (the biggest footgun), async Server Component data fetching, waterfalls, streaming, after(), and metadata. Use when writing or reviewing Next.js — app/ routes, 'use client'/'use server', Server Components, route handlers, next.config, caching, revalidation — or a React Server Components question. Defers to writing-ts for pure TypeScript.

Orchestrate parallel autonomous implementation using worker subagents in isolated worktrees. Reads the implementation plan from project docs/memory and dispatches workers.

Analyze test coverage gaps (unit + E2E), create missing tests using worker subagents, then run and fix all tests. Uses Explore subagent for analysis and Worker subagents for parallel test creation.