Engineering Discipline

Production-grade AI coding discipline. Prevents the top 4 AI coding sins: acting without thinking, over-engineering, collateral damage, and vague execution. Based on Karpathy's 4 rules + battle-tested additions (3-layer consistency checks, anti-rationalization, verification loops, surgical diffs). Works with Claude Code, Cursor, Copilot, and any AI coding assistant.

When to Use

Apply this skill to every coding session. It's not a task-specific tool — it's a permanent behavior modifier for AI coding assistants.

Especially critical when:

• Working on production codebases (>1000 lines)
• Making changes that touch multiple files or components
• The AI assistant starts "suggesting improvements" you didn't ask for
• You notice the AI making assumptions about your intent

The 4 Foundational Rules (Karpathy)

Rule 1: Think Before Coding

Problem: AI acts on assumptions, not understanding.

Before writing any code:

1. If the requirement is ambiguous → ask, don't guess
2. If there are multiple valid approaches → list them with tradeoffs
3. If the request seems wrong → push back with reasoning
4. If you're uncertain about scope → confirm before touching files

❌ Bad: "I'll refactor this module while fixing the bug" ✅ Good: "The bug is in line 42. I see the surrounding code could be cleaner, but you only asked me to fix the bug. Should I also refactor?"

Rule 2: Simplicity First

Problem: AI defaults to over-abstraction.

• 10 lines > 100 lines for the same result
• No feature creep — only build what was asked
• No premature abstraction — don't add interfaces "just in case"
• Litmus test: would a senior engineer say "this is too complex"? → rewrite

❌ Bad: Adding a factory pattern, three interfaces, and a config layer for a simple utility function ✅ Good: One function, clear name, no unnecessary indirection

Rule 3: Surgical Changes

Problem: AI makes "drive-by" edits to code it wasn't asked to touch.

• Fix the bug, only the bug
• Don't reformat adjacent code
• Don't update comments you weren't asked about
• Don't change variable names in unrelated functions
• Every changed line must trace back to the user's specific request

❌ Bad: "While fixing the auth bug, I also cleaned up the logging format and renamed some variables" ✅ Good: 3 lines changed, all in the auth function, all directly related to the bug

Rule 4: Goal-Driven Execution

Problem: Vague instructions lead to vague results.

Instead of telling the AI how to do something, give it a success criterion:

❌ "Fix the login bug" ✅ "Write a test that reproduces the login timeout on slow networks, then make it pass"

❌ "Improve the API" ✅ "Response time for /api/users must be under 200ms for 1000 concurrent requests"

The AI iterates better toward measurable goals than fuzzy directions.

💡 Why This Way: LLMs are natural iterators. Given a clear target, they'll loop (generate → test → adjust) until they hit it. Given a vague goal, they'll generate once, declare victory, and move on.

Battle-Tested Additions (Beyond Karpathy)

A1: Three-Layer Consistency Check

After any change, verify alignment across layers:

Layer 1 — Naming: env vars, DB columns, API paths, config keys must match across all files Layer 2 — Business: design docs ↔ code ↔ UI ↔ API responses must tell the same story Layer 3 — Database: migrations ordered correctly, FK references valid, types match TS interfaces

Run the relevant layer after each change. Run all three on major releases.

A2: Anti-Rationalization

Never trust the AI's "I think this looks correct."

• "I read the code" ≠ verified → run it
• "It should work" ≠ confirmed → test it
• "I wrote it, so it's right" = rationalization → verify independently

A3: Verification Loop

For every change type, define a verification action:

Changed	Verify by
Code/script	Execute it
Config	Restart + confirm effect
Generated file	Check content (wc -l, grep, diff)
API call	Check return value
UI change	Visual diff before/after

A4: Pre-Change Snapshot

Before modifying any file:

1. Record current state (grep key content, or screenshot)
2. Make the change
3. Diff to confirm only intended parts changed
4. If unintended changes found → revert and redo surgically

A5: Context Hygiene

AI context windows are finite. Polluted context → degraded output.

• Trim tool outputs (pipe to head -30, don't dump 500 lines)
• Checkpoint progress to files during long tasks
• Don't let the AI "remember" — make it read files

Integration

Claude Code (CLAUDE.md)

Add to your project's CLAUDE.md:

# Engineering Discipline Rules
[paste the 4 rules + additions above]

Cursor (.cursor/rules)

Add to .cursor/rules/engineering-discipline.md

Any AI Coding Tool

These rules work as system prompts, project instructions, or conversation primers for any LLM-based coding assistant.

Related Skills

• trinity-harness — Full agent harness with Challenge + Execute + Compound layers
• self-improving-agent — Continuous learning from mistakes
• skill-creator — Create new skills from workflows

Feedback

• If useful: clawhub star engineering-discipline
• Issues: https://github.com/clawhub/engineering-discipline