chenlu-hung

chenlu-hung/my-skills

Skill Claude CodeCodex

Formalizes the Anthropic-recommended architect/implementer split — the strongest model (Fable 5, this session) grills the user via the grill-me skill, designs the architecture, and writes it to a self-contained doc; an implementer subagent (default Opus, overridable to sonnet/haiku via argument) then implements…

chenlu-hung/my-skills

Skill Claude CodeCodex

Ask the ChatGPT desktop app a question from the command line and read the answer on stdout, through the chatgpt-ask CLI — by default in a throwaway temporary chat, or, for research collaboration, inside a ChatGPT project or an existing saved thread, plus reading a whole thread's transcript and listing a project's…

chenlu-hung/my-skills

Skill Claude Code

Delegates coding tasks to external agent CLIs as if they were subagents — Codex (ChatGPT sub), Gemini (Antigravity agy), and DeepSeek (opencode, free) — each through its own subscription/sign-in CLI, not an API key. Workers get write access and can run in isolated git worktrees for parallel work; this Claude Code…

chenlu-hung/my-skills

Skill Claude CodeCodex

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree — then write the agreed plan as a task list and, once approved, work through it in dependency order, dispatching the parallelizable tasks to model subagents. Use when user wants to…

chenlu-hung/my-skills

Skill Claude Code

A context-transfer process for AI coding sessions that creates a short handoff document for another session or agent.

chenlu-hung/my-skills

Skill Claude Code

Convenes a multi-model "council" to answer a question, then synthesizes a single best answer — inspired by Karpathy's llm-council. Each member runs through its own subscription/sign-in CLI, not an API key: ChatGPT (the desktop app, via chatgpt-ask), Gemini (Antigravity agy), and Claude (claude -p), with Codex (codex…

chenlu-hung/my-skills

Skill Claude Code

Builds and maintains a compact, on-demand "project map" so coding agents remember a codebase's structure and spend fewer tokens during maintenance. Combines deterministic ctags indexing and an import-based module dependency graph with short LLM-written module summaries, committed to the repo. Use when the user says…

chenlu-hung/my-skills

Skill Claude Code

Sends your own plan, diff, or design out to independent external models (Codex, Gemini, optionally a fresh Claude) to find what you MISSED — the failure mode of a single strong model working alone. Reviewers read the real repo read-only, so they can catch omissions your prompt never mentioned. Four modes: plan (before…

chenlu-hung/my-skills

Skill Claude Code

Picks the right writing register per output segment, automatically. Chat answers stay dense without going telegraphic; text destined for a file reads as deliberately written rather than as compressed chat or formulaic AI prose; code and structured data are left byte-exact. Replaces the separate caveman (compression)…