Find agent skills by outcome

Behavioral guidelines to reduce common LLM coding mistakes

derived from Andrej Karpathy's observations on LLM coding pitfalls.

- State your assumptions explicitly. If uncertain

ask.

Tradeoff: These guidelines bias toward caution over speed. For trivial tasks

use judgment.

description: Behavioral guidelines to reduce common LLM coding mistakes. Use when writing

reviewing

2. Communication — clear

confident

karpathy-guidelines

1. Content Quality — specific

relevant

- Interview type: [e.g.

HR screening / Technical / C-level / panel]

- After 8–10 questions

end the interview naturally and tell me: We'll be in touch. Type ANALYZE when you're ready for feedback.

- Mix question types: behavioral (STAR)

technical

- After I answer each question

give a brief 1-line reaction (like a real interviewer would — neutral

Mockup Interview using Gemini Live

${job_title} at [COMPANY TYPE/NAME].

2. Base typography variables (--font-body

--font-heading

5. Utility classes or direct styles for special cases (caps

tabular numbers

- NEVER pair two visually similar fonts (e.g.

two geometric sans-serifs)

- Max 3 heading levels. If you need H4+

restructure the content.

- MUST use unitless line-height values (1.3-1.45)

never px or em

- Bold not italic for headings. Subtle size increases (1.2-1.5x steps

not 2x jumps)

- Line length drives breakpoints

not the other way around

- Paragraph spacing via margin-bottom equal to one line-height

no first-line indent for web

- font-variant-numeric: tabular-nums on data/number columns

oldstyle-nums for prose

- Use clamp() for fluid sizing: `clamp(1rem

0.9rem + 0.5vw

- NEVER use vw units alone (breaks user zoom

accessibility violation)

- Free quality options: Source Serif

IBM Plex