claude-toolkit

Run the announcement round for a finished project: attribution check, directory submissions, launch-post drafts in Mike's voice. Drafts only — never posts. Trigger: /announce <project>, 'announce', 'launch round', 'tell people about'.

Novel ignition ritual for story-bible: 3-line 'previously on' + one pulling question. Body-double for getting into writing state, zero obligation. Trigger: /ignite, 'ignite', 'novel time', 'story time'.

Read/post Twitter/X via `bird` CLI: tweets, threads, search, timelines, bookmarks, follows, lists, media.

Drive VISIBLE Chrome with real profile/logins (CDP :9222). Screenshots, JS eval, interactive element picking, cookies. Frontend testing, authed flows, anything user should see. Headless text-only reading → Browser* tools. Linux + macOS.

Read before git commits.

Compact current conversation into handoff doc for another agent.

Run and monitor GPU experiments with `jl` CLI on JarvisLabs.ai.

Empirically tune GGUF model's llama-swap/llama.cpp config for multi-GPU rig — card placement, KV-cache quant, context length, batch sizes, prompt-cache settings. Benchmarks each dial one at a time, verifies with real load. Checks Hugging Face model card for sampling/template. Modes: "bench" (max tok/s, benchmark-maxxing), "agentic" (max context + safe VRAM headroom for coding agents). Use when user wants to tune, optimize, speed up, or fit local LLM (llama.cpp / llama-server / llama-swap / GGUF) on GPUs — "tune a model for speed", "max out context for a model", "what settings for this model", "optimize my llama-swap config", "dial in tok/s".

LocalMaxxing API for local-LLM benchmark + eval leaderboards. Trigger: submitting benchmarks, pushing tok/s or eval scores.

Browser automation with Playwright. Auto-detects dev servers, writes scripts to /tmp. Testing websites, forms, login flows, screenshots, responsive design, any browser task.

Whole-repo audit for over-engineering. Like ponytail-review but scans entire codebase, not diff: ranked list of what to delete, simplify, or replace with stdlib/native. Trigger: "audit this codebase", "audit for over-engineering", "what can I delete from this repo", "find bloat", "ponytail-audit", /ponytail-audit. One-shot report, no fixes.

Code review hunting over-engineering only. Finds what to delete: reinvented stdlib, unneeded deps, speculative abstractions, dead flexibility. One line per finding: location, what to cut, what replaces it. Trigger: "review for over-engineering", "what can we delete", "is this over-engineered", "simplify review", /ponytail-review. Complements correctness review — complexity only.

Forces laziest solution that works — simplest, shortest, most minimal. Senior dev seen everything: question whether task needs to exist (YAGNI), stdlib before custom code, native platform before deps, one line before fifty. Levels: lite, full (default), ultra. Trigger: "ponytail", "be lazy", "lazy mode", "simplest solution", "minimal solution", "yagni", "do less", "shortest path", or complaints about over-engineering, bloat, boilerplate, unnecessary deps.

Extremely strict maintainability review: abstraction quality, giant files, spaghetti-condition growth. Trigger: thermo-nuclear code quality review, thermonuclear review, deep code quality audit, harsh maintainability review.

Apply Strunk's writing rules to ANY human-read prose — docs, commit messages, error messages, explanations, reports, UI text. Makes writing clearer, stronger, more professional.

Ultra-compressed communication mode. Cuts tokens ~75%. Levels: lite, full, ultra, wenyan-lite/full/ultra. Trigger: /caveman, "be brief".

Write codebase docs: README, architecture, API, comments. Project structure and getting-started focus.

React/Preact frontend conventions: TanStack Query for server state, 8-section components, no TypeScript, no barrel files.

Find architecture improvements: consolidate modules, improve testability and AI navigation. Trigger: improve architecture, refactor.

Build throwaway prototype — terminal app for state/logic, or UI variations. Trigger: prototype this, let me play.

Diagnose hard bugs, regressions, and test failures via feedback-loop discipline and root-cause first. Trigger: diagnose, debug, broken.

Cloudflare APIs: DNS, Workers, Tunnels, zones. Trigger: Cloudflare, DNS records, manage domains.

Create distinctive frontend interfaces: components, pages, dashboards, React, HTML/CSS. Polished code, no generic AI aesthetics.

TypeScript/full-stack review: type safety (branded, unions, end-to-end), real tests over mocks, OpenTelemetry, right abstractions.

TDD with red-green-refactor. Build features or fix bugs. Trigger: red-green-refactor, integration tests, test-first.

Create isolated git worktrees for feature work. Smart selection, safety verification. Use: isolation, before implementation.

Screenshot and analyze visual composition: golden ratio, rule of thirds, spacing, hierarchy. Verify UI quality before presenting.