English | 日本語

cat-skills

The daily toolkit for developers and designers who build with AI agents.

Why cat-skills?

Some skills sharpen the agent — prodev, elevate, and eye load concrete standards and expert lenses that fix how the model thinks, builds, and judges for the entire session.

Some give you superpowers for building — recon gives you senior-engineer research instincts before you write a line of code, architect gives you a structured build methodology that doesn't lose the plot, and liquid-cat-physics takes all of it and runs the build autonomously while you're away.

And some are just useful every day — eat turns any YouTube video, article, or podcast into usable knowledge in seconds, notebook keeps your project context alive across sessions, xray gives you social intelligence on X.

This is what catcatcat uses every day to design, build, and ship. 14 skills, each independent, each doing one thing well.

Installation

Claude Code

npx skills add catcatcatstudio/cat-skills

Cursor

cp -r .cursor your-project/

Note: Cursor skills require setup:

1. Switch to Nightly channel in Cursor Settings → Beta
2. Enable Agent Skills in Cursor Settings → Rules

Learn more about Cursor skills

Gemini CLI

cp -r .gemini your-project/

Note: Gemini CLI skills require setup:

1. Install preview version: npm i -g @google/gemini-cli@preview
2. Run /settings and enable "Skills"
3. Run /skills list to verify installation

Learn more about Gemini CLI skills

Codex CLI

cp -r .codex/* ~/.codex/

OpenCode

cp -r .opencode your-project/

Kiro

cp -r .kiro your-project/

Pi

cp -r .pi your-project/

Trae

# International
cp -r .trae/skills/* ~/.trae/skills/

# China
cp -r .trae-cn/skills/* ~/.trae-cn/skills/

Note: Trae has two versions with different config directories:

• International: ~/.trae/skills/
• China: ~/.trae-cn/skills/

After copying, restart Trae IDE to activate the skills.

Then type any command: /eat, /architect, /eye, /prodev, etc.

What's Inside

Think — judgment, standards, dissent

Build — plan, research, execute

Ship — test, persist, maintain

Orchestrate

Skills

Elevate — Expert Elevation

Shifts the model from compliant executor to critical expert advisor. Auto-detects the domain, adopts the mindset of a top-tier practitioner (not generic "senior engineer" — specific, like "someone who's built Figma, Linear, and Notion"), and produces ranked proposals.

Will push back if the work is already strong — proposing zero changes is a valid output. Works across all domains: UI/UX, code, copywriting, architecture, strategy, branding.

Source: skills/elevate/SKILL.md

Prodev — Engineering Standard

Loads a full field manual that changes how the model evaluates every decision, every line, every assumption. Not "try harder" — a concrete behavioral shift for the rest of the session.

Includes a self-review checklist that runs before every commit. If any answer is "no," the code doesn't ship.

Source: skills/prodev/SKILL.md

Adversary — Structured Dissent

Pressure-tests decisions through adversarial analysis. Auto-detects the domain (architecture, strategy, marketing, design), builds the strongest case FOR the current approach, then spawns an adversary subagent to steel-man a concrete alternative at full strength.

Produces a decision brief with load-bearing assumptions, the strongest counter-argument, the single question that resolves the disagreement, and a mandatory verdict (Hold / Adjust / Reconsider / Uncertain). Short-circuits when there's no credible alternative — doesn't manufacture doubt.

Source: skills/adversary/SKILL.md

Eye — Design Judgment

Design critique through the lens of a design director with taste — not checklists, not scores, not rubrics. Scales from micro-reviews (a single element) to full design crits (an entire app).

Includes a design knowledge base that calibrates LLM blind spots: confusing competence with quality, rule-following without context, AI aesthetic bias, describing instead of judging, spacing blindness, and confusing minimalism with emptiness.

Source: skills/eye/ · README

Architect — Staged Build

Drives software projects through a structured lifecycle: spike, design, plan, build. Prevents the #1 AI coding failure — losing context on a big plan and missing pieces.

Source: skills/architect/SKILL.md

Recon — Pre-Build Intelligence

Surveys the landscape before you write any code. Adopts an expert consultant mindset — not "I googled best practices" but "here are the 5 decisions that will make or break this project, ranked by how hard they are to fix later."

Generates expert questions before searching, uses parallel research agents, prioritizes primary sources (engineering blogs, postmortems, issue trackers) over listicles. Outputs an opinionated summary with the top decisions, what surprised the expert, and the biggest unknown.

Source: skills/recon/SKILL.md

Eat — Knowledge Extraction

Pulls transferable knowledge from any URL or content — YouTube, Instagram, TikTok, X videos, podcasts, articles, X threads, PDFs.

Full pipeline: download → transcribe (Whisper/Groq) → frame extraction → visual assessment → knowledge synthesis.

Strips noise (ads, filler, self-promotion). Preserves signal (frameworks, methods, specific numbers, practitioner honesty).

Requires: yt-dlp, ffmpeg. Transcription: whisper (local) or GROQ_API_KEY. Optional: defuddle, X_BEARER_TOKEN, browser cookies for social platforms. See setup instructions.

Source: skills/eat/SKILL.md

Xray — X/Twitter Content Intelligence

Graph-based X intelligence. Build a social graph of accounts you engage with, track, and study — then scout for reply opportunities, pulse-check topic lanes, track competitors, mirror your own performance, and prospect for clients.

Includes setup interview, 8 lane presets, full-archive search, and cost tracking.

Source: skills/xray/ · README

Fortify — Testing Infrastructure

Detects your stack, installs the complete testing ecosystem, audits code for untested critical paths, writes thorough tests weighted toward error paths, then runs them and verifies they catch real bugs through mutation testing.

Source: skills/fortify/SKILL.md

Vet — Codebase Investigation

Opens a project, maps the architecture, investigates every layer for real problems — structural issues, logic bugs, security gaps, reliability holes, performance traps, stubs, AI slop, dead code — then builds a prioritized fix plan and executes collaboratively. Carries its own engineering standard (internalized from prodev) so it operates at senior-engineer judgment without needing a separate lens loaded.

Six scan categories: Structural (architecture, coupling, god files, legacy paths), Correctness (logic bugs, stubs, edge cases, race conditions), Security (injection, auth gaps, data exposure), Reliability (error handling, defensive code, resource leaks), Performance (N+1, unbounded ops, hot paths), Hygiene (dead code, type drift, AI comment noise).

Source: skills/vet/SKILL.md

Scorched Earth — Exhaustive Analysis

When surface-level isn't enough and the cost of missing something is higher than the cost of being slow. Scorched Earth is obsessive thoroughness with a paper trail — every file, every function, every claim examined, verified, and logged. It's deliberately too thorough: it will take a while, and it will surface things you didn't ask about. That's the point.

Three modes, picked from what you point it at:

Pairs naturally with Vet: reach for Vet when you want problems found and fixed with judgment; reach for Scorched Earth when certainty that nothing was missed matters more than speed.

Source: skills/scorched-earth/SKILL.md

Notebook — Project Notes

Prevents context loss and reasoning loops. Maintains a trail of decisions, failures, and lessons learned.

Source: skills/notebook/SKILL.md

Memento — Context Handoff

Saves unsaved session knowledge to notebook, then produces a compact orientation block for fresh chats. Also includes an auto-compaction safety net — hooks that preserve and re-inject conversation context when Claude Code's context window fills up.

Source: skills/memento/ · README

Liquid Cat Physics — Autonomous Deep-Work Loop

Turns Claude into its own project manager. Reads project state, applies an expert lens to decide what to work on, enforces senior-engineer execution standards on every line of code, gates every action through a confidence check, and persists everything. Loops every 10 minutes.

Two embedded lenses drive quality: the elevate lens picks the smartest next move (not just the next TODO), and the prodev standard ensures it's built right — correct layer, verified APIs, understood blast radius, no stubs, self-reviewed before every commit. Includes a three-tier confidence gate (GREEN/YELLOW/RED), two-strike anti-thrashing, automatic checkpoints with coverage assessment, and memento auto for context survival.

Source: skills/liquid-cat-physics/ · README

How They Compose

The skills work independently, but some are designed to work together:

elevate + prodev + eye — Three lenses for three concerns. Elevate decides what to work on. Prodev governs how to build it. Eye judges whether it looks right. Each sharpens a different axis.

architect + notebook — Architect writes to notebook format natively. Notebook adds proactive saves, context recovery, and lesson tracking on top.

liquid-cat-physics — The orchestrator. Embeds elevate and prodev as reference lenses, uses notebook for persistence, memento for compaction survival. One command to start, then walk away.

Supported Tools

Claude Code · Cursor · Gemini CLI · Codex CLI · OpenCode · Kiro · Pi · Trae

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

License

MIT © catcatcat