ai-startup-skills
AI 제품을 만드는 스타트업을 위한 7단계 에이전트 스킬: 검증 → 비즈플랜 → 정책 → 로드쇼 → 고객 확보 → 가격 책정 → 등록
저장소 정보는 아직 수집하지 않았습니다.
아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.
Collection of my agent skills (and skills that I found good or regular use)
저장소 README에서 가져옴
README 전체를 불러오는 중입니다.
README 전문 보기 ↗A personal collection of agent skills — both custom-authored and cherry-picked from upstream repos — installable as a single package.
📚 Read the docs for the full guide, conventions, and skill index.
# Install all skills bundled in this repo into your project's
# .agents/skills/ directory.
npx skills@latest add daviddwlee84/agent-skills/skills
Note The trailing
/skillsmatters. Without it,npx skillswill look in.agents/skills/of the upstream repo, which contains a different layout. Theskills/suffix points the installer at theskills/local/,skills/owned/, andskills/vendor/trees.
This is the recommended cross-agent install path. Project installs place the
canonical skill content under .agents/skills/, which Codex reads directly;
Claude Code receives its corresponding .claude/skills/ links.
Heads-up — upstream picker bug.
npx skills' interactive multiselect glitches once the skill list is long enough to scroll: options render duplicated and the selection state gets corrupted (vercel-labs/skills#969). This repo ships enough skills to trigger it. Until it's fixed upstream, skip the picker:# Install everything, non-interactively (no picker) npx skills@latest add daviddwlee84/agent-skills/skills --yes # List names from a disposable directory. In an existing agent-managed # project, [email protected] may reconcile .agents/skills even with --list. scratch="$(mktemp -d)" (cd "$scratch" && npx skills@latest add daviddwlee84/agent-skills/skills --list) rm -rf "$scratch" # Install one skill by name. npx skills@latest add daviddwlee84/agent-skills/skills --skill clash-proxy-api
A local checkout can expose the same category catalog through either native plugin CLI:
git clone https://github.com/daviddwlee84/agent-skills.git
# Claude Code
claude plugin marketplace add ./agent-skills/skills
claude plugin install version-control@daviddwlee84-skills
# Codex (verified with 0.147.0)
codex plugin marketplace add ./agent-skills/skills
codex plugin add version-control@daviddwlee84-skills
Both native routes install a whole category plugin (version-control in this
example), not an arbitrary single skill. Codex reads the existing Claude-format
marketplace and generates its .codex-plugin/plugin.json adapter inside the
plugin cache; this repository does not need to maintain that second manifest.
Keep using npx skills when you want one skill or a cross-agent installation.
Pass the checkout's skills/ directory, not the repository root: the catalog
intentionally lives under skills/.claude-plugin/ for the npx subpath above,
and both native CLIs require a supported manifest at the marketplace source
root.
The quantitative-finance marketplace group contains quantatitive-factor-researcher,
vectorbt, and nautilus-trader; MLflow, DVC, and batch notebooks remain in 04-ml-workflow.
skills/local/ — custom-authored skills.
project-knowledge-harness — TODO + backlog + pitfalls structure with init / kanban / promote / add-todo / sweep-inbox toolkit. (docs)quantatitive-factor-researcher — Python quant-research persona. (docs)vectorbt — OSS/PRO library development, XDG documentation sync, migration guidance, and optional official MCP setup. (docs)nautilus-trader — NautilusTrader development: detects the v1 Cython vs v2 Rust/PyO3 generation, syncs version-matched docs/examples from git tags into an XDG cache, and gates sandbox/live trading. (docs)skill-author — Author new skills following agentskills.io best practices; ships new-skill.sh scaffolder and lint-skill.sh linter. (docs)The grouped picker UI of npx skills@latest add daviddwlee84/agent-skills/skills
is driven by skills/.claude-plugin/marketplace.json.
Each plugin entry defines one category and explicitly lists its member skills;
anything not listed there falls under Other automatically. The manifest is
the source of truth for the current category set rather than a duplicated list
in this README.
# Validate the manifest (paths exist, no duplicates, name not reserved,
# on-disk skills covered).
make marketplace
# Exercise Claude Code + Codex native add/list/install in isolated state.
make native-marketplace-smoke
The manifest lives under skills/, not at the repo root, because the
install command's /skills subpath makes the CLI read
<repo>/skills/.claude-plugin/marketplace.json. See
npx skills metadata model
for the full mechanism, the kebabToTitle group naming rule, reserved
marketplace names, and the marketplace.json vs plugin.json distinction.
This repo dogfoods project-knowledge-harness on itself:
TODO.md — priority/effort-tagged backlog.backlog/ — research notes for P? items (when populated).backlog/inbox.md — quick-capture inbox; sweep into TODO.md later.pitfalls/ — symptom-grep-able knowledge base of past traps (when populated).# Validate TODO.md and render kanban-style board
make kanban
# Quick add structured TODO entry
./scripts/add-todo.sh --priority P3 --effort M --title "..." --description "..."
# Move an active TODO to ## Done in same commit as the implementation
./scripts/promote-todo.sh --title "<substring>" --summary "<shipped summary>"
# Triage backlog/inbox.md interactively
./scripts/sweep-inbox.sh
Full workflow: Project memory.
# Add a vendored skill (auto-syncs)
./scripts/add-vendor.sh owner/repo/path/to/skill
# or via Makefile
make add-vendor SOURCE=owner/repo/path/to/skill
# Group into a series subdir (skills/vendor/<series>/<name>/)
./scripts/add-vendor.sh --series fullstack-nextjs vercel/vercel-plugin/skills/nextjs
# Sync all vendored skills from upstream
make sync
# Check for upstream updates (dry-run)
make sync-check
Full workflow: Adding vendor skills.
Dependencies: gh (GitHub CLI, authenticated) and yq (YAML processor).
The docs site is built with MkDocs Material + mkdocs-llmstxt + mkdocs-copy-to-llm.
uv sync --extra docs
make docs-serve # http://127.0.0.1:8000/
make docs-build # produces ./site/
Deployed automatically to GitHub Pages on push to main via .github/workflows/docs.yml.
If you want to apply the same docs stack to your own project, see Downstream docs stack recipe.
Use the bundled scaffolder (recommended — seeds the agentskills.io best-practice template and fans out discovery symlinks for non-universal agents):
# Auto-detects scope: LOCAL (this repo) / PROJECT (another git repo) / GLOBAL (~/).
bash skills/local/skill-author/scripts/new-skill.sh <skill-name>
# Override scope explicitly when intent isn't obvious:
bash skills/local/skill-author/scripts/new-skill.sh --project <skill-name>
bash skills/local/skill-author/scripts/new-skill.sh --global <skill-name>
bash skills/local/skill-author/scripts/lint-skill.sh skills/local/<skill-name>
The scaffolder mirrors npx skills add's placement contract — canonical
content goes under .agents/skills/<name>/ (read directly by Cursor /
Codex / OpenCode / Warp / Gemini-CLI / Copilot), with relative symlinks
fanned out to .claude/skills/ and any other already-present
non-universal agent dir. In this repo (LOCAL mode) the canonical lives
in skills/local/<name>/ and the symlinks point back at it. See
Creating local skills
for the full scope table.
Or, just the SKILL.md skeleton via the upstream init (no symlinks, no
references/ / scripts/ / assets/):
cd skills/local
npx skills@latest init [skill-name]
The full curated index of upstream skill collections, MCP servers, and domain-specific hubs lives in the docs site:
📚 Catalog — external skill collections, per-domain hubs (Finance, Quant Research, AI/ML Research, Web & Fullstack, Knowledge Work, Agent Harness), a docs-only curiosity shelf, and an MCP wiki.
Highlights:
vendored / deferred / skipped / evaluated / wishlist status per entry.anthropics/financial-services, RKiding/Awesome-finance-skills, the Financial Datasets MCP, and more.verifiable-surfaces--help--dry-run--print-config12-factor-agent-design-review — Design or evidence-review production LLM applications using HumanLayer's 12-Factor Agents principles; covers LLM/code boundaries, prompts, context, typed tools, durable state, pause/resume, humans, retries, and replay. (docs)openui-direct-provider — Build OpenUI generative UI while keeping inference on your own provider (OpenAI/Anthropic/Google/OpenAI-compatible/local). It never pushes the paid Gateway unless you ask, and it repairs invalid output with your own model. Provider-neutral companion to vendored openui. (docs)demo-evidence — Capture acceptance evidence (screenshots, terminal/screen recordings, HTTP logs) for a just-built feature into a gitignored .evidence/ bundle keyed to git branch/commit + coding-agent session, for async "Demos over diffs" review. Lifecycle: new-bundle.sh → capture.sh <web|term|http|screen> → finalize.sh. (docs)mkdocs-site-bootstrap — Bootstrap a MkDocs Material site + GitHub Pages deploy; consent-gated via .skills/preferences.yaml; ongoing add-docs-page.sh helper. (docs)zensical-site-bootstrap — English Zensical pilot with safe initialization, locked docs-only tooling, native LLM exports, and build-only CI; not a MkDocs/i18n migration. (docs)marimo-batch-mlflow — Opinionated fork of upstream marimo-batch: Tyro CLI (dataclass or Pydantic) + MLflow tracking + live mlflow-widgets chart, dual-mode (mo.app_meta().mode == "script") UI/CLI from one notebook. (docs)dvc-ml-workflow — DVC (treeverse/dvc) pipelines + queued experiments with metrics auto-bound to ephemeral commits; ships init-dvc-project.sh, queue-helper.sh (with grid cartesian-product enqueue), lint-dvcyaml.sh. (docs)mlflow-tracking — Generic MLflow skill: SQLite + mlflow ui for solo, vendored PostgreSQL + MinIO docker-compose stack for teams; covers LLM tracing, model registry (aliases), and autologging across all officially-supported frameworks. (docs)agent-history-hygiene — Commit SpecStory chat transcripts + .claude/plans/*.md alongside feature diffs; bootstrap pre-commit + gitleaks + secret-redactor into new repos; enforce rotate-first, rewrite-last leak remediation (never reflexive git push --force). (docs)git-workflow — One opinionated, scale-aware git workflow (solo → team): Conventional Commits (English-only), linear history via pull --rebase + fast-forward, tiered main-vs-dev-vs-PR (GitHub Flow), worktrees for parallel agents, SemVer release tags, and pre-merge secret hygiene (defers to agent-history-hygiene). Ships branch-status.sh (classify branches) + check-commit-msg.sh (Conventional Commits validator). (docs)fastapi-ai-patterns — Production FastAPI patterns + gotchas for AI/ML/LLM serving: the def/async def decision table, event-loop/lifespan/response_model/BOLA traps up front, and 8 references covering all 10 chapters. Inspired by FastAPI for AI Engineers (original wording). (docs)fastapi-ai-scaffold — Generate a runnable, production-shaped FastAPI AI service via new-fastapi-ai-service.sh (bash 3.2, --dry-run/JSON): a 44-file clean-architecture skeleton with lifespan-loaded model, JWT, SSE LLM gateway, guardrails, /health+/ready, tests, Docker. Verified green out of the box. (docs)fastapi-ai-interview-prep — 100 self-written FastAPI/AI-engineering interview Q&A across 10 topics, each with a senior-signal rubric, plus a stdlib quiz.py mock-interview CLI (--random/--topic/--answer/--json). (docs)raycast-extension-dev — Build, verify, and ship Raycast extensions: the four-stage gate (tsc --noEmit -> dev-check -> ray lint -> ray build, because ray build bundles with esbuild and never typechecks), the launchd PATH trap that a dev terminal structurally hides, transport-seam design for CLI-wrapping extensions, ack-before-apply reconcile timing, MenuBarExtra constraints, and check-store-readiness.sh for the store requirements ray lint passes over (verified: it exits 0 with an empty metadata/). (docs)clash-proxy-api — Discover and drive a running Clash/mihomo proxy via its external-controller API: clash_api.py (stdlib-only) for status/node/mode/TUN/reload/connections/latency with robust controller discovery (not just 9090 — handles Verge Rev 9097), plus clash_sysproxy.sh to toggle the OS system proxy (macOS networksetup / GNOME gsettings) and references for enabling the API per client (mihomo CLI, Clash Verge Rev, ClashX). (docs)python-project-best-practice — Modern Python project conventions for the agentic era: uv + src/ layout, PEP 735 [dependency-groups], Tyro CLIs (completion is built in — no shtab), loguru at entry points only, and ruff/type/pytest gates behind a Justfile. Ships new-python-project.py (six additive profiles: minimal/cli/lib/api/ml/rust, --dry-run, JSON out) and audit-python-project.py (26 read-only checks + an ordered migration plan, no --fix), plus an AGENTS.md CLI block regenerated from --help and gated in CI. All six profiles verified green end-to-end. (docs)skills/owned/ — first-party skills maintained in another project, synced through vendor.yaml with collection: owned. Edit the canonical source there; this repo distributes the copy. See the owned workflow.
go-cli-tui — From awesome-lazy-tools; build Go CLIs, Lazygit-inspired dashboards, and guided wizards with Charm, arrow/Vim navigation, XDG preferences, and real terminal verification. (docs)userscript-development — Maintained in Tampermonkey-Scripts: SPA/GM/runtime gotchas, Markdown extraction, Chromium/Firefox fixtures, real-manager installation and native-extension comparison. (docs)skills/vendor/ — third-party skills synced from upstream via vendor.yaml. Do not edit these locally; make sync will clobber changes. Skills can be flat (skills/vendor/<name>/) or grouped into a series (skills/vendor/<series>/<name>/).
marimo-notebook — from marimo-team/skills; general marimo authoring conventions. (docs)streamlit-to-marimo — from marimo-team/skills; convert Streamlit apps to marimo notebooks. (docs)anywidget (frontmatter name: anywidget-generator) — from marimo-team/skills; generate anywidget components for marimo notebooks. (docs)skill-creator — from anthropics/skills; evaluate, benchmark, and optimize skill trigger descriptions. Complements local skill-author. (docs)herdr — official Herdr terminal/agent-control skill. For machines with the binary installed, prefer herdr --skill so guidance matches that exact release. (docs)security-audit — from cloudflare/security-audit-skill; defensive, source-first security review. Guidance mode by default; the full six-phase audit (sandboxed validation, independent verification, schema-checked findings.json) runs only on explicit request. (docs)openui — official OpenUI skill from thesysdev/skills; generative UI with OpenUI Lang, component libraries, and Agent Interface. Gateway-first: it recommends the paid OpenUI Gateway unless you ask to call your provider directly. (docs)fullstack-nextjs — Next.js (App Router) + Supabase (Postgres) + shadcn/ui + Tailwind CSS + design/testing skills, all from official orgs. See the series overview.
nextjs — Next.js App Router expert from vercel/vercel-pluginshadcn — shadcn/ui CLI + components from vercel/vercel-pluginreact-best-practices — TSX reviewer (70+ rules) from vercel/vercel-pluginvercel-storage — Blob/Edge Config + Supabase/Prisma integration from vercel/vercel-pluginsupabase — full Supabase product surface from supabase/agent-skillssupabase-postgres-best-practices — Postgres perf rules from supabase/agent-skillsweb-design-guidelines — UI audit reviewer from vercel-labs/agent-skillsfrontend-design — anti-AI-slop aesthetics from anthropics/skillswebapp-testing — Playwright toolkit from anthropics/skills