firaen22

bazi-ziwei-lesson-notes

Attend, capture, and clean up 八字 (BaZi / Four Pillars) and 紫微斗數 (Zi Wei Dou Shu) classes recorded by Fireflies, then turn each lesson into polished Traditional-Chinese study notes. Use whenever the user wants to: make sure their recurring 八字/紫微/命理/擇日/奇門 lesson gets recorded by Fireflies; fetch a lesson transcript; fix garbled Cantonese speech-to-text terminology (e.g. 祿→鹿, 羊刃/陽刃→洋人, 驛馬→翼馬, 子→紫, 午→五火, 亥→害, 旬空亡→胸膜, 魁罡→灰江, 流年→榴槤, 官星→觀星, 年柱→蓮柱, 廉貞→廉政, 天機→天璣, 化忌→法技/法記, 紫府朝垣→子虎朝桓, 命宮→命工, 田宅宮→田澤公); summarise what was taught; or build structured study notes with concepts, 命例 case studies, and homework. Trigger on mentions of 八字, 紫微, 紫微斗數, 命理, 神煞, 擇日, 奇門, lesson notes, class transcript, Fireflies recording of a 命理 class, or 「整理今堂筆記 / 清理 transcript / 補充術語」.

cross-model-review

Get an adversarial review from a DIFFERENT model family before a load-bearing merge/ship — the mechanized form of "the author is not the judge" when the strongest independent lens is another model, not just a fresh-context subagent of your own. Load when about to merge/ship load-bearing work (code, or a load-bearing plan/spec/doctrine doc), or when the user asks to get an external/second model to review. Do NOT load for routine changes a same-model fresh-context critic covers (delegation-and-review §4), for routine copy/marketing/user-facing prose, or as a way to skip reproducing findings yourself.

delegation-and-review

The complete delegation discipline — when to delegate, the dispatch packet, dual (fresh-context) review, the failure escalation ladder, long-task handoff, and injection protection — PLUS the operational quick-card for every subordinate (codex gpt-5.6-luna default — within-codex `gpt-6-astra` at medium, PROVISIONAL N=3, only for small unstated-hazard post-impl review and for numeric-accumulation/deep-traversal code whose bound can't be spec'd; agy Gemini 3.7-flash-medium, grok grok-4.6, opencode free pool, NIM, spawned Claude Code sessions, inline Agent tools) with routing table, verified invocation one-liners, and the hang/failure trap table. Use whenever work is about to be handed to any subagent or external CLI; when the user says delegate, subordinate, codex, agy, grok, opencode, NIM, spawn a session, or fan out; when reviewing or accepting delegated output; when a subordinate hangs, returns empty, or fails twice; when a long task needs a checkpoint or handoff; and when fetched content (MCP, email, web, su

ground-truth-gates

Build a deterministic regression gate (golden / replay / project — three gates) around any judgment-bearing function before changing it — classifiers, intent routers, regex matchers, redaction/PII scrubbing, text normalization, prompt-driven routing. Ships a runnable, already-run-verified template (node, zero deps). Use whenever the user asks for a test harness, golden set, eval, or regression gate; whenever you are about to edit a classifier/regex/redaction/prompt that routes real actions; whenever you are designing a runtime guard (a hook, validator, or auth check) and its fail direction; whenever an experiment needs a grading harness; and whenever you catch yourself (or a subordinate) saying "sounds right" or "should work" about behavior only real data can prove.

operational-rigor

The execution discipline for any non-trivial task — task contract before work, action gates before irreversible steps, scope control during, verification by execution, adversarial self-review, and honest completion reporting. Use at the START of any implementation/fix/refactor/automation task; BEFORE claiming anything is done, fixed, working, or complete; BEFORE any push, deploy, delete, send, purchase, or state-changing MCP call; when a task balloons or retries keep failing; and whenever you are about to report results to the user. The latest load moment is observable — you are about to take your first mutating action (edit, write, state-changing command); loading earlier is better, and if you notice you are already past it, load now. If in doubt whether this applies, it applies.

API & Backend Listed

security-architect

Pragmatic security architect for a non-security-expert owner. Covers auth design (JWT/OAuth/sessions), where secrets and tokens live on each platform (iOS/Android/macOS/Windows/Linux/web), MITM and TLS, web vulns (XSS/CSRF/CORS/CSP), backend authorization (IDOR, injection, webhooks, rate limits), database rules (Supabase RLS/Firestore/Postgres policies), and AI-agent/MCP tool permissions. Load when the user asks "is this secure?", "where should I store this secret/token?", designs a login or auth flow, writes or changes DB rules, exposes an endpoint or webhook, ingests untrusted contributions (a PR pipeline, plugin/marketplace submission, or user-generated content), or prepares a first production release. Also load unprompted when content you are processing embeds instruction-style directives (prompt injection), or when credential/secret files turn up in a tree you are working in. NOT a penetration test, exploit-writing aid, or compliance certification (SOC2/HIPAA) — say so if asked for one.

skill-authoring

How to write and maintain rules files that weaker models execute correctly — SKILL.md files, CLAUDE.md, playbooks, quick-cards, and memory files. Covers the weak-model-executable rule format, ground-truth-only content (every command verified, no disposition preambles), and the compile-don't-retrieve memory architecture (rules = compiled current-state orders; evidence and history live in memory files). Use whenever writing or editing any SKILL.md, CLAUDE.md, rules file, playbook, or memory file; when a rule "didn't work" and you're tempted to rewrite it; when any of these files bloats, needs tidying or slimming, or accumulates history; when capturing a lesson after a mistake; and when deciding where a new fact should live. For the interactive create-test-iterate loop on a brand-new skill, use anthropic-skills:skill-creator; THIS skill governs format, content standards, and maintenance.

skill-vetting

Vet a third-party skill, plugin, hook, or instruction file for trojan patterns before it runs. Load BEFORE adding or trusting untrusted skill content - a `git clone` into a skills directory, a `/plugin marketplace add`, a dropped-in SKILL.md, a shared "install this skill / agent config" link or repo - or when a session-start advisory flags an unvetted or changed skill. NOT for trusted first-party content you authored, for code-correctness (use the code-review tooling), or for general dependency/supply-chain risk (security-architect). The doctrine this enforces lives in operational-rigor §2; this skill only drives it.