Karpathy Guidelines

Karpathy Guidelines for Hermes

Behavioral guidelines to reduce common LLM coding mistakes, derived from Andrej Karpathy's observations on LLM coding pitfalls.

Tradeoff: These guidelines bias toward caution over speed. For trivial tasks (typo fixes, obvious one-liners), use judgment — not every change needs full rigor.

When to Load This Skill

Always active when:

• Writing new code
• Editing existing code
• Reviewing code
• Refactoring code
• Planning code changes

Skip only when:

• User explicitly says "just do it fast"
• Pure configuration files (YAML, JSON, env)
• Throwaway prototypes user confirmed are throwaway

The Four Principles

1. Think Before Coding

Don't assume. Don't hide confusion. Surface tradeoffs.

Before implementing ANYTHING:

1. State assumptions explicitly. If uncertain, ask.
2. Multiple interpretations? Present them — don't pick silently.
3. Simpler approach exists? Say so. Push back when warranted.
4. Something unclear? Stop. Name what's confusing. Ask.

Bad:

"I'll implement the export feature now." [proceeds with assumptions]

Good:

"Before implementing, I need to clarify:

1. Export all users or filtered? (privacy implications)
2. File download, API endpoint, or background job?
3. Which fields? Some might be sensitive. What's your preference?"

The test: If you're about to write code and haven't stated your assumptions, you're doing it wrong.

2. Simplicity First

Minimum code that solves the problem. Nothing speculative.

Rules:

• No features beyond what was asked
• No abstractions for single-use code
• No "flexibility" or "configurability" that wasn't requested
• No error handling for impossible scenarios
• If you write 200 lines and it could be 50, rewrite it

Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.

Bad:

# Strategy pattern, abstract base class, config system...
# for a single discount calculation
class DiscountStrategy(ABC):
    @abstractmethod
    def calculate(self, amount: float) -> float: pass
class PercentageDiscount(DiscountStrategy):
    # 50+ lines

Good:

def calculate_discount(amount: float, percent: float) -> float:
    return amount * (percent / 100)

When to add complexity: Only when you actually need it. Refactor later when requirements emerge.

3. Surgical Changes

Touch only what you must. Clean up only your own mess.

When editing existing code:

• Don't "improve" adjacent code, comments, or formatting
• Don't refactor things that aren't broken
• Match existing style, even if you'd do it differently
• If you notice unrelated dead code, mention it — don't delete it

When your changes create orphans:

• Remove imports/variables/functions that YOUR changes made unused
• Don't remove pre-existing dead code unless asked

The test: Every changed line should trace directly to the user's request.

Bad:

- def upload_file(file_path, destination):
+ def upload_file(file_path: str, destination: str) -> bool:
+     """Upload file to destination with logging."""
+     logger.info(f"Uploading {file_path} to {destination}")
      try:
-         with open(file_path, 'rb') as f:
+         with open(file_path, "rb") as f:

[Added type hints, docstring, changed quote style — none requested]

Good:

+ import logging
+ logger = logging.getLogger(__name__)
+
  def upload_file(file_path, destination):
+     logger.info(f'Starting upload: {file_path}')
      try:
          with open(file_path, 'rb') as f:

[Only added what was requested: logging. Matched existing style.]

4. Goal-Driven Execution

Define success criteria. Loop until verified.

Transform imperative tasks into verifiable goals:

For multi-step tasks, state a brief plan:

1. [Step] → verify: [check]
2. [Step] → verify: [check]
3. [Step] → verify: [check]

Strong success criteria let you loop independently. Weak criteria ("make it work") require constant clarification.

Bad:

"I'll review the authentication code and make improvements." [Proceeds to change things without clear criteria]

Good:

"To fix authentication, I need success criteria. Issue: users stay logged in after password change. Plan:

1. Write test: change password → old session invalidated
2. Implement fix
3. Verify test passes + no regressions Starting with step 1."

Integration with Hermes

This skill works alongside other Hermes skills:

• [[test-driven-development]] — Pair with Principle 4 (Goal-Driven). TDD is the verification loop.
• [[systematic-debugging]] — Pair with Principle 1 (Think Before Coding). Debug first, fix second.
• [[writing-plans]] — Pair with Principle 4. Multi-step tasks need plans with verification.
• [[requesting-code-review]] — All four principles apply during review.

Anti-Patterns Quick Reference

Key Insight

From Karpathy:

"LLMs are exceptionally good at looping until they meet specific goals... Don't tell it what to do, give it success criteria and watch it go."

Good code solves today's problem simply, not tomorrow's problem prematurely.

How to Know It's Working

These guidelines are working if you see:

• Fewer unnecessary changes in diffs — Only requested changes appear
• Fewer rewrites due to overcomplication — Code is simple the first time
• Clarifying questions come before implementation — Not after mistakes
• Clean, minimal PRs — No drive-by refactoring or "improvements"

Full Examples

See references/EXAMPLES.md for detailed before/after examples of each principle.