Towards AI

ANTHROPIC RELEASED THE FULL GUIDE ON HOW TO USE CLAUDE CODE.

It's an agent.. Stop using it like autocomplete.

Inside:
→ CLAUDE.md
→ Plan mode
→ Verification
→ Subagents

Save this for your next AI engineering build.

vibe coding gets you a prototype. a spec gets you a product

it's literally one file. what you're building, what done means, what the agent can't touch

weak spec + strong model = beautiful code for the wrong thing

Interviewed another candidate today.

I asked him what a system prompt is.

He replied:
Instructions that set the model's behavior, separate from user messages.

Then I asked about few-shot prompting.

He said:
Showing examples in the prompt so the model learns the pattern in-context.

Then I asked a follow-up:
Your prompt works perfectly in the playground.

In production it fails 20% of the time.

Same model.

What's different?

What would your answer be?

you don't have an eval problem, you have a logging problem

can't evaluate a failure you never captured. input, retrieved context, output, tool calls, cost

if you can't replay yesterday's bad answer, those aren't evals. that's vibes with a dashboard