Free AI skills from Agent Dev Thailand.

ฝรั่งแจกพร้อมท์ฟรีกัน ทำไมคนไทยต้องซื้อพร้อมท์

Install all skills (Claude Code / Cursor / Codex / Windsurf + more):

npx skills add somnus0x/agt-skill-pack

Install one skill:

curl -sL https://raw.githubusercontent.com/somnus0x/agt-skill-pack/main/install.sh | bash -s factory-review

Install all via script (Claude Code only):

curl -sL https://raw.githubusercontent.com/somnus0x/agt-skill-pack/main/install.sh | bash

Clone and go (Claude Code / Codex / Cursor):

git clone https://github.com/somnus0x/agt-skill-pack.git
cd agt-skill-pack
# Open in your AI tool and start talking

Any AI chat (Claude.ai / ChatGPT / Gemini): Open the skill file, copy the prompt, paste as your first message.

"ผมให้ AI ตัวอื่น review ก่อน deploy ทุกครั้ง"

Use a different AI model to review anything your primary AI produced. Same-model self-review has blind spots — a different model catches what yours missed.

5 review modes: Content, Spec/Architecture, Code, Fact-Check, Quick Review

Try: "review this draft" · "red team this spec" · "sanity check this code"

📄 skills/factory-review/SKILL.md

"ก่อน build feature ถามตัวเอง 5 ข้อนี้ก่อน"

5-filter evaluation for product decisions. Runs features through who-suffers, replaces-what, core-loop, cost-of-being-right, and ship-without-it filters. Kills bad ideas before you waste a sprint.

Try: "should we build this?" · "taste check" · "what would you cut?"

📄 skills/product-taste/SKILL.md

"คุณเคยกลับไปดูมั้ยว่า decision เดือนก่อนยังถูกอยู่?"

Re-evaluate past decisions when assumptions change. Scores decay as Fresh, Aging, or Stale. Because a choice that was correct in January may be wrong in March.

Try: "is this still the right call?" · "decay check" · "revisit our DB decision"

📄 skills/decision-decay/SKILL.md

"ให้ AI ไล่ถามว่าทำยัง"

Personal commitment tracker with escalating confrontation. Three levels: firm reminder → confrontational → no mercy. Tracks streaks, skips, and the pattern of avoidance you're pretending doesn't exist.

Try: "nag me about gym" · "accountability check" · "did I actually do it?"

📄 skills/accountability-nag/SKILL.md

"UI คุณดู AI-generated อยู่มั้ย"

Checklist to identify AI-generated UI patterns. Scores 15 tells across color, layout, visuals, typography, and motion. If your score is 6+, the UI screams "AI made this."

Try: "does this look AI?" · "slop check" · "design audit"

📄 skills/ai-slop-detection/SKILL.md

"เลิกถามว่า UI สวยมั้ย"

The judgment layer above the slop checklist. Start from a defined design identity, not a blank screen — generic is a starting-point problem. Four lenses to enforce it: squint test, remove until it breaks, copy test, emotional register. Plus typography/color/spacing/motion references.

Try: "design review" · "taste check" · "does this look good?"

📄 skills/design-taste/SKILL.md

"ช่วยจัด workflow ให้หน่อย"

8-question interview → Workflow Map → AI setup plan → role-specific prompts. Builds your AI workflow from scratch based on your actual work.

Try: "help me set up AI for my work" · "ช่วยจัด workflow ให้หน่อย"

📄 skills/workflow-scout/SKILL.md

"save ลิงก์ไว้ 200 อัน อ่านจริง 3 อัน"

Drop any URL into your AI — get a 3-5 line summary, category, and relevance score. Know what's worth reading without reading everything. Works with articles, tweets, YouTube, GitHub repos, PDFs.

Try: "triage this link" · "summarize this URL" · "what's this about?"

📄 skills/link-triage/SKILL.md

"ไม่ได้ก๊อปโค้ด แต่ขโมยไอเดีย"

Daily scan of GitHub trending filtered for your stack. AI picks 3-5 repos with ideas worth stealing — patterns, architectures, and techniques you can adapt for your project. Set it as a cron and it runs every morning.

Try: "steal digest" · "what's trending?" · "anything worth stealing?"

📄 skills/steal-digest/SKILL.md

"ก่อนเขียน ดูก่อนว่าใครเขียนอะไรไปแล้ว"

Research Twitter/X before creating content. Find what's performing, what angles are saturated, and where the content gaps are. Write something new, not something redundant.

Try: "scout this topic" · "content gaps for AI tools" · "what are people saying about X?"

📄 skills/x-collect/SKILL.md

"AI ลืมทุกอย่าง? เพราะคุณไม่เคยสร้างทางให้มันจำ"

Two drop-in prompts that give your AI persistent memory. Long-term: FTS5 search over past sessions — find old discussions in milliseconds. Short-term: hot cache file so AI knows current project state at session start. No cloud, no API keys.

Try: "set up AI memory" · "search past sessions" · "AI doesn't remember anything"

📄 skills/ai-memory/SKILL.md

"ผมรัน AI 3 terminal พร้อมกัน แล้วมันแก้ทับกันเองจนพัง"

Run 2-3 agents in parallel and they overwrite each other's files. The fix is how you'd fix a colliding human team: one agent owns one domain, the rules live in a contract file, you stop being the bottleneck. Drop-in prompt to map your repo into bounded domains + the per-agent CLAUDE.md boundary block. Concepts from Domain-Driven Design + Team Topologies, dropped onto agents.

Try: "agents keep colliding" · "split my repo into domains" · "manage agents like a team"

📄 skills/agent-boundaries/SKILL.md

"เปิด Claude Code ครั้งแรกกับโปรเจค React — มันเขียน code ถูก แต่ไม่เข้ากับโปรเจค"

One-time setup that detects your React/Next.js stack and generates CLAUDE.md rules matching your actual project. Component conventions, import rules, styling approach, data fetching patterns, TypeScript config — all detected from your existing code, not generic best practices.

4-step workflow: Detect → Generate → Guardrail → Verify

Try: "setup frontend" · "configure for react" · "frontend onboarding"

📄 skills/frontend-setup/SKILL.md

"AI สรุปตัวเลขให้ผิดประจำ — เพราะให้มันไปหา data เองตอน gen"

When AI fetches live data while it's generating, it guesses from half-loaded pages and gives you wrong numbers. The fix: split the two beats — cron caches every source to JSON first, then one prompt reads the files. The AI never touches the network, so it stops guessing.

The pattern: cron (fetch) → data/*.json → one prompt (read + summarize) → brief Built-in guard: "do not guess numbers; missing = n/a" — so a dead feed writes n/a, not a hallucination

Try: "setup data brief" · "morning brief pipeline" · "stop AI guessing numbers"

📄 skills/market-brief-setup/SKILL.md

"ผมเอา game theory มาใส่ใน AI review — แล้วมันเริ่มทำงานต่างออกไป"

Tit-for-tat cooperation tracking for AI reviews. Your reviewer remembers what it flagged last round, adjusts depth based on your track record, and catches recurring issues by name. Based on the strategy that won Robert Axelrod's 1984 programming tournament.

3 review depths: Light (high trust) · Standard · Deep Audit (low trust) Auto-correction loop: 3 iterations, model rotation, self-fixing drafts

Try: "axelrod review" · "trust score" · "why does it keep flagging the same thing?"

📄 skills/axelrod-review/SKILL.md

"เราแค่ถาม ไม่ได้สั่ง แต่มันแก้ไปแล้ว 3 ไฟล์"

The AI reads exploratory messages as commands and builds on a guess. Hold Up is the brake: when it catches discuss-mode cues ("i think", "should we", a question with no imperative verb) right as it's about to edit/commit/deploy, it stops and asks ONE tight question instead of sprinting. Pairs with grill-me (the invoked deep-interview) — Hold Up is the automatic front-door reflex that decides whether you even need one.

The give: a 2-line CLAUDE.md rule that makes the brake fire automatically.

Try: "/hold-up" · "just talk, don't action this" · "let's think first"

📄 skills/hold-up/SKILL.md

Each skill works on all major AI coding tools. Clone once, use anywhere.

Available inside any skill session:

• SKIP — jump to next section
• EXPAND [topic] — go deeper
• DONE — summarize with current data
• REDO — start over

Skills should be free. Inspired by dot-skill.