30-second
wins.save these!

Wrap pasted material in tags like <document>…</document> or triple quotes. The model can tell your instructions apart from the content you want it to work on.

Add "If you're not sure, say so rather than guessing." It's a simple line that cuts down on confident nonsense.

Put your role, writing style and pet hates into ChatGPT's custom instructions (or a Claude project) so you stop repeating yourself in every chat.

Long chats collect stale context. When answers get worse, ask for a summary of decisions so far, open a new chat and paste it in.

Ask the coding agent for a plan first and review it before any file changes. Fixing a plan costs seconds; unpicking a bad diff costs an afternoon.

Have the AI write a failing test for the bug, then fix the code until it passes. A test is a goal it can check itself.

Use /clear in Claude Code when you switch to an unrelated task. Leftover context from the last job quietly confuses the next one.

Put framework versions, folder conventions and "don't add new dependencies" in Cursor's project rules so every chat starts with them.

Commit (or stash) before letting an agent make big changes. git diff shows you exactly what it did, and a reset undoes it.

Don't write "no text" in the prompt. Use the --no text parameter instead, because mentioning a word often adds it.

Found an image whose style you love? Use it as a style reference with --sref so a whole campaign looks like it belongs together.

"Golden hour side light", "soft overcast" or "neon rim light" change an image more than adding adjectives about the subject.

In n8n, pin the output of a trigger node while building. You can rerun later steps without sending yourself 40 test emails.

Let automations draft emails and posts, but route them to a human for approval until you trust the output. Then remove the gate.

Every production automation needs a failure alert (email or Slack). Silent failures are worse than no automation.

If every API call repeats the same long instructions or document, turn on prompt caching. Repeated input gets much cheaper and faster.

Before changing a prompt or model in production, rerun a small fixed set of real questions. Vibes are not a test suite.

In RAG, split documents on headings or paragraphs with a little overlap, so an answer that spans a boundary isn't lost.

Ollama unloads idle models after a few minutes. Set keep_alive on requests if you want instant replies during a session.

Paste a scary spreadsheet formula and ask "explain this step by step, then suggest a simpler version." Great for inherited files.

When summarising a meeting transcript, ask for decisions, owners and deadlines as a table. Nobody reads paragraph minutes.

When an AI gives sources, open at least the ones you'll rely on. Links can be real but not say what the summary claims.

Paste three pieces of writing you're proud of and ask the AI to describe the voice as rules. Reuse that guide in every marketing prompt.