The agent skills, system prompt, and CLI toolkit I use every day with Claude Code, Codex, and other coding agents.
Landing page: https://cskwork.github.io/skills/ · 한국어 요약은 아래에 있습니다.
Pick the skills you want, and which agents to install them on:
npx skills@latest add cskwork/skillsList them without installing:
npx skills@latest add cskwork/skills --listInstall the system prompt (back up your existing ~/.agents/AGENTS.md first):
mkdir -p ~/.agents ~/.claude ~/.codex
curl -fsSL https://raw.githubusercontent.com/cskwork/skills/main/system-prompt/AGENTS.md -o ~/.agents/AGENTS.md
ln -sfn ~/.agents/AGENTS.md ~/.claude/CLAUDE.md
ln -sfn ~/.agents/AGENTS.md ~/.codex/AGENTS.mdInstall the local voice skill:
npx skills@latest add cskwork/skills --skill caveman -gThe included system prompt does not set a reply style. To make Caveman your default, add to your prompt: "Use the caveman skill by default for user replies; explicit style requests override this default." Say stop caveman or normal mode to disable it for the session. attention-kind remains available on request. This installs the voice skill only; it does not install the Caveman proxy, routing, or telemetry.
system-prompt/AGENTS.md is the operating contract I give every coding agent: agree on outcome and scope, ground decisions in evidence, verify, and report plainly. Source of truth: cskwork/THE-SYSTEM-PROMPT.
| Skill | What it does | Source |
|---|---|---|
| sdlc-kit | Four-phase SDLC loop (intent → plan → build → evidence) with human gates, adversarial review, and scored, routed proof. | cskwork/sdlc-kit |
| verify-app | Build and run the app to prove a change works end to end. For agents without a built-in verify, such as Codex. | this repo |
| bughunt | Find and fix interaction bugs with seeded personas, invariant oracles, and failure replay. App profiles for games, CRUD apps, AI chat, and media; Godot, native iOS, and web adapters. | cskwork/bughunt-skill |
| quality-loop | Product quality bar and improvement loop: scores UI/UX, function, consumer trust, and engineering against hard floors, gates release candidates with an independent judge, ratchets every fix in a ledger, and reports quality as a % (target >= 90%) with the reason for each lost point. | cskwork/quality-loop-skill |
| qa | /qa launches five parallel QA subagents (frontend users acting out real usage, mistakes, and unexpected actions, or backend analysts checking APIs and data against cited domain intent), reproduces every finding, and returns one report with fix proposals. Default Sonnet 5.5 medium; model, effort, and agent count are arguments. |
cskwork/qa-skill |
| db-intelligence | One skill for PostgreSQL, MySQL, SQLite, and MongoDB: detect the engine, connect credential-safe, read schema, read before writing, never write to prod. | cskwork/pi-setup-public |
| sql-walkthrough | Explain SQL/batch logic to a non-expert: numbered source lines, the real rows they read, per-line arithmetic, a results table, and an executed counter-example. | cskwork/sql-walkthrough |
| improve-codebase-architecture | Find deepening opportunities, show them as an HTML report, then grill through the one you pick. | mattpocock/skills |
| grill-with-docs | Relentless design interview that sharpens a plan while keeping GLOSSARY.md and ADRs updated as decisions settle. |
mattpocock/skills |
| domain-modeling | Actively build and sharpen a project's domain model in GLOSSARY.md and docs/adr/. |
mattpocock/skills |
| research | Investigate a question against primary sources in a background agent and record cited findings in Markdown. | mattpocock/skills |
| prototype | Build a throwaway prototype to answer a design question. | mattpocock/skills |
| retro | Run a retrospective on a coding session. | mattpocock/skills |
| jk | Drive Jenkins from the terminal with the jk CLI: jobs, runs, logs, artifacts, credentials. |
avivsinai/jenkins-cli |
| swiftui-pro | Review SwiftUI code for modern API use, data flow, navigation, accessibility, and performance, with before/after fixes. Pairs with mobile-ui-ux. | twostraws/SwiftUI-Agent-Skill |
| simless | Default way to test iOS apps: render SwiftUI screens as accessibility trees, hot-reload views, and run unit tests on the Mac without a simulator; the simulator is kept for the final pass. | abacusai/simless |
| Skill | What it does | Source |
|---|---|---|
| grilling | Interview the user in frontier rounds until a plan or design tree is settled with nothing silently assumed. | mattpocock/skills |
| teach | Teach a new skill or concept over multiple sessions in a stateful workspace (mission, glossary, resources, learning records). | mattpocock/skills |
| caveman | Terse reply style: answer first, preserve technical facts and exact payloads. Local voice skill only. | JuliusBrussee/caveman |
| attention-kind | Answer in the ADHD-friendly Attention-kind style: one-sentence bottom line first, arrow-marked bold points, short by default, expanding only on what is vital while preserving numbers, conditions, and warnings. | alexgreensh/attention-span |
| fluent-korean | Write clear, fluent Korean: keep particles and endings, end sentences with a predicate, use plain words instead of metaphors, avoid em-dashes. Converted from an output style to a skill. | snflkd/fluent-korean |
| handoff | Compact the conversation into a handoff document for another agent. | mattpocock/skills |
| writing-for-agents | How to write skills, AGENTS.md, and CLAUDE.md for agents. | mattpocock/skills |
| ponytail | Lazy-senior-dev mode: the simplest, shortest solution that actually works. | DietrichGebert/ponytail |
| work-status | Report progress on the current request in numbers: items done, in progress, todo, and blocked, evidence for each done item, and an ETA for the rest. | this repo |
| confluence-explain-doc | Explain a bug's cause and fix in Confluence so a first-time reader gets it: real code captures with red/green/blue line highlights, one section per problem, how problems connect, git blame plus the original reason for each cause, details folded. Project specifics live in a separate domain profile. | this repo |
| Skill | What it does | Source |
|---|---|---|
| impeccable | Design, critique, audit, and polish frontend interfaces. | pbakaus/impeccable |
| gpt-image-2 | Generate and edit images with ChatGPT's current image model through your ChatGPT plan and the Codex CLI. Model-agnostic: follows whatever model Codex uses. | agentspace-so/agent-skills (modified) |
| diagram-design | 44 diagram and chart types (architecture, flowchart, sequence, state, ER, timeline, Sankey, and more) as self-contained HTML/SVG/PNG, with Mermaid, .drawio, and .excalidraw import. | cathrynlavery/diagram-design |
| motion-graphics | Code-driven motion-graphics videos: product launch videos (Apple/Google keynote style) from real captures of your running app, and vertical YouTube Shorts, Reels, and TikTok. | cskwork/motion-graphics-skill |
| mobile-ui-ux | Measurable UI/UX rules for iPhone apps, mobile web, and mobile games (touch targets, safe areas, contrast, Korean line breaks, HUD and combat readability), plus a 7-category scoring rubric, an independent-judge prompt, a pre-ship checklist, and per-project overrides. Pairs with impeccable. | written in this repository |
| image-first-ui | Image-first UI: three UX-distinct interface options generated as images through your ChatGPT plan; the one you pick becomes the reference the real build must match. Adds production sizing, stack choice, a domain pattern list, and a release gate with an independent judge. Admin, web app, landing page, or game screens. | written in this repository (image scripts from gpt-image-2) |
| Skill | What it does | Source |
|---|---|---|
| call-agent | Delegate a bounded task to a peer CLI: Codex, Claude, Antigravity, Kiro, NotebookLM, ChatGPT Pro. | cskwork/call-agent |
| update-all | Update all coding agent harnesses (Codex, Claude Code, pi, Antigravity CLI) in one run with before/after version reporting. | cskwork/update-all-skill |
| sol | Orchestrate app work through GPT-6.1 Sol workers: sliced briefs, one git worktree per parallel slice, sandbox-aware preflight, session-resume retry, and orchestrator-run acceptance checks. | cskwork/sol-skill |
These skills ship with a tool or a machine setup, so get them from their own home instead of copying them here.
| Skill | What it does | Install |
|---|---|---|
| verify | Claude Code's own /verify: builds and runs your app to confirm a change does what it should. |
Comes with Claude Code (v2.1.200+). Type /verify. |
| ego-browser | Browser for agents that shares your logged-in sessions (ego-lite). | Install the ego lite app, or npx skills add citrolabs/ego-lite |
| teams-native-chat | Send text and clickable links through the signed-in Teams web client's native chat service, without Microsoft Graph or webhooks. | Install from the linked standalone repository. |
| hindsight | Shared long-term memory for Claude Code, Codex, Hermes, and Pi (hindsight-agent-setup). | curl -fsSL https://raw.githubusercontent.com/cskwork/hindsight-agent-setup/main/bootstrap.sh | bash |
| Skill | What it does | Access |
|---|---|---|
| cswap-auth | Authenticate Claude Code accounts through Ego Lite, renew OAuth credentials, register with cswap, and verify account health. | Private repository; access required. Not included in the public install. |
The CLIs around the agents: account switching, status line, Google Workspace, Firebase, Jenkins, agent runtimes, shell output compression (rtk), and my Pi setup. See toolkit/README.md.
My own files are MIT (LICENSE). Vendored skills keep their own license file in their folder; see THIRD_PARTY.md for sources and the exact commits copied.
매일 쓰는 에이전트 스킬, 시스템 프롬프트(AGENTS.md), CLI 도구 모음입니다. npx skills@latest add cskwork/skills로 원하는 스킬만 골라 설치합니다. 다른 사람이 만든 스킬은 원본 라이선스와 출처를 그대로 유지합니다.