common-sense-skill

Common Sense Skill

One file. 9 questions. Stops your AI agent from rm -rf-ing your repo, spamming clients, or buying €5K hardware when €30 solves it.

A drop-in skill that turns any AI agent from a blind executor into a thoughtful operator — by forcing a practical-judgment check before every consequential action.

TL;DR

git clone https://github.com/cfilipemt/common-sense-skill ~/.claude/skills/common-sense
# Same drop for Codex, Cursor, OpenClaw, Aider — into their skills dir.

Auto-fires before any consequential action. No config.

👉 See 10 real production incidents this prevents 🐦 Launch copy / share snippets

---

Intelligence without common sense is dangerous. Power without judgment is destruction. Speed without clarity creates chaos.

---

Why This Exists

LLM agents are fast, capable, and too willing to act.

They will:

• Run rm -rf when asked to "clean up"
• Push to production at 4 AM on a Friday
• Send 50 cold emails to a single recipient because the loop didn't break
• Recommend buying €5,000 hardware when €30 solves the problem
• Confidently invent API endpoints
• Optimize for being right instead of getting the right result

Common sense is the missing layer.

This skill installs that layer.

---

What It Does

Forces the agent to ask 9 practical questions before any consequential action:

1. Does this make sense?
2. Is this safe?
3. Is this necessary?
4. Is this the smartest path?
5. What are the consequences?
6. What would a highly responsible human do here?
7. Is there a simpler and better solution?
8. Does this create unnecessary risk?
9. Would this decision still make sense tomorrow, next month, next year?

Plus a 4-question decision filter for the big moves:

• Would I do this if my own money depended on it?
• Would I do this if my family depended on it?
• Would I defend this decision publicly?
• Would I repeat this decision 100 times?

Any "no" — don't proceed.

---

The 9 Principles

#	Principle	Core idea
1	Practical Judgment	Simplicity beats complexity
2	Consequence Awareness	Every action has ripple effects
3	Human Reality Check	Emotion + context matter
4	Context Intelligence	Same rule may fail in different situations
5	Prevention Over Repair	Avoid disaster, don't just fix it
6	Responsibility	Act like an owner, not a tool
7	Truth Over Ego	Accuracy > being right
8	Resource Awareness	Time, money, energy are limited
9	Silent Signal Detection	Read between the lines

Full protocol in SKILL.md.

---

Before vs After

Without the skill

User:   "Clean up the old logs"
Agent:  $ rm -rf /var/log/* 
        ✓ Done
Result: Server unrecoverable. Carlos hates you.

With the skill

User:   "Clean up the old logs"
Agent:  Common sense check:
        - Is this safe? Wildcard delete on /var/log/ kills active syslog + auth logs
        - Smartest path? Rotate + archive, don't nuke
        - What does a responsible sysadmin do? logrotate, not rm -rf
        
        Proposed: archive logs >30 days to /backup, delete archives >90 days.
        Confirm before run?
Result: Safe cleanup. Trust grows.

---

How It Triggers

The skill auto-loads when the user says:

• /common-sense
• "common sense"
• "does this make sense"
• "is this practical"
• "real world"
• "would a normal person"
• "is this sensible"

Or it auto-applies silently before any consequential action:

• Production deploys
• File edits in critical paths
• Outbound messages (email, WhatsApp, SMS)
• Recommendations (purchases, tools, vendors)
• Alerts, notifications, escalations
• Schema migrations
• Public-facing changes

---

Install

Claude Code

mkdir -p ~/.claude/skills/common-sense
curl -fsSL https://raw.githubusercontent.com/cfilipemt/common-sense-skill/main/SKILL.md \
  -o ~/.claude/skills/common-sense/SKILL.md

Codex (Anthropic / OpenAI Codex CLI)

mkdir -p ~/.codex/skills/common-sense
curl -fsSL https://raw.githubusercontent.com/cfilipemt/common-sense-skill/main/SKILL.md \
  -o ~/.codex/skills/common-sense/SKILL.md

OpenClaw

mkdir -p ~/.openclaw/skills/common-sense
curl -fsSL https://raw.githubusercontent.com/cfilipemt/common-sense-skill/main/SKILL.md \
  -o ~/.openclaw/skills/common-sense/SKILL.md

Cursor / Aider / Continue / any agent framework

Drop SKILL.md into your framework's system-prompt loader or skill directory. The content is plain Markdown — load it as instructions, project rules, or a tool description.

One-liner (all three)

for D in ~/.claude/skills/common-sense ~/.codex/skills/common-sense ~/.openclaw/skills/common-sense; do
  mkdir -p "$D"
  curl -fsSL https://raw.githubusercontent.com/cfilipemt/common-sense-skill/main/SKILL.md -o "$D/SKILL.md"
done

---

Production Examples

Real cases this skill has caught in production:

Scenario	What agent wanted to do	Common-sense flag
Family + 1-bed apartment	Show as match	Excluded — too small for family
Retiree + 4th floor no lift	Show as match	-25 score, flag accessibility
Buyer rotation script	Email same client 5 times	Rate-limit, dedupe by name
WhatsApp voice reply	Reply with text	Mirror modality — reply with voice
Database migration	Run during peak hours	Defer to off-peak window
New domain deploy	Skip status-page entry	Force entry per "always list every domain" rule
403 from upstream API	Retry forever	Backoff + restart watchdog

---

Output Pattern

Silent application when the check passes — no narration:

(agent proceeds normally)

Visible flag when the check changes the recommendation:

🦞 Common sense check: X looks reasonable because Y. Watch out for Z.

⚠️ Common sense flag: this might break X. Recommend Y instead.

Never overrides explicit user instruction — escalates with the concern.

---

Integration Matrix

Framework	Tested	Auto-trigger	Manual /common-sense
Claude Code	✅	✅	✅
OpenAI Codex CLI	✅	✅	✅
OpenClaw	✅	✅	✅
Cursor	➖	➖	✅ (via Rules)
Aider	➖	➖	✅ (via .aider-instructions)
Continue.dev	➖	➖	✅ (via system prompt)
Generic LangChain	➖	manual	✅ (load as system prompt)

Tested = used in real production stack. ➖ = should work, untested by maintainer.

---

Mental Model

┌────────────────────┐
│   User Request     │
└─────────┬──────────┘
          ▼
┌────────────────────┐
│  Agent Drafts Plan │
└─────────┬──────────┘
          ▼
┌────────────────────────────────────┐
│  COMMON SENSE CHECKPOINT           │
│  • 9 questions                     │
│  • 9 principles                    │
│  • 4-question decision filter      │
└────────┬───────────────┬───────────┘
         │ PASS          │ FAIL
         ▼               ▼
   ┌──────────┐    ┌──────────────┐
   │  Action  │    │  Block +     │
   │ Proceeds │    │  Flag user   │
   └──────────┘    └──────────────┘

---

Roadmap

• More language translations (PT, IT, FR, DE, ES)
• Domain-specific extensions (medical, finance, legal)
• Tested integrations for Cursor, Aider, Continue
• Optional --strict mode (always flag visibly)
• Community examples library

---

Related Work

The phrase "common sense" appears in AI research with a different meaning. Worth understanding the distinction before you compare:

Project	Focus	Layer
ConceptNet (MIT, 2.9k⭐)	Factual world knowledge graph — "apples are fruit", "rain makes things wet"	Static facts
ConceptNet Numberbatch (1.3k⭐)	Pretrained word embeddings derived from the above	Static facts
Open Mind Common Sense	The umbrella research project that started it all (MIT Media Lab, 1999)	Static facts
This skill	A judgment protocol applied at the moment of action	Runtime decision

ConceptNet answers "what does the world look like?" This skill answers "should I actually do this?"

Both matter. Neither replaces the other. An agent with a great world model can still confidently rm -rf your repo — because knowing facts isn't the same as knowing when to stop and ask.

If you're building agents and want both:

• Use ConceptNet (or embeddings) for retrieval and grounding
• Use this skill at the decision gate, before irreversible actions

---

Contributing

PRs welcome. Keep the protocol short. Common sense is short by design.

If you add a new principle: argue why it can't be folded into an existing one.

---

License

MIT — use it everywhere.

---

Author

Built by Carlos Filipe. Battle-tested across multiple production AI agents.

---

⭐ Star the repo if this saved you from a rm -rf moment.