Guides from the night desk
Animated shorts about coding with AI. Each guide is the fix from one episode. Contact: whyshouldaicare@gmail.com
The Claude Code survival kit
Every lesson from the series, as something you can paste. One rule per disaster we already had, so you don't have to.
Make your AI tell you why it fails
Assistants are trained to agree. Ask for the flaws instead of a verdict.
Make Claude Code format every file, always
Instructions are suggestions. A hook runs every time, whether Claude remembers or not.
Same question, different answers: what to do
Models pick each word with a bit of randomness (temperature). Two runs can disagree; neither is automatically true.
Check a package exists before you install it
AI assistants invent package names. Attackers register them (slopsquatting).
Stop is not undo: rewind in Claude Code
Esc stops Claude. It does not put your files back. Checkpoints do.
Embeddings, with code you can run
An embedding turns text into a point on a map of meaning. Close points mean similar meaning.
One format, many videos, no lookalikes
Our own checklist after nine of eighteen videos came out as twins.
Safe permissions instead of --dangerously-skip-permissions
Allow the boring commands, deny the destructive ones, get asked about the rest.
When to make Claude think (and when not)
Thinking is the model talking to itself before answering. You pay for those tokens.
Split one giant PR into reviewable ones
Nobody reads 3,000 lines. They approve them.
Plan mode: fix the plan, not the code
Shift+Tab twice in Claude Code: Claude reads and proposes steps, and changes nothing until you approve.
Teach Claude your style without fine-tuning
Most "train it on my code" is one good CLAUDE.md and one example.