For critical information, ask two different models (Claude and GPT, say) the same question independently, and compare before relying on either.

Before using any reference or quote AI cites (book title, article, study), verify it actually exists before citing it publicly.

Train yourself to notice AI speaks with the same linguistic confidence whether 100% correct or entirely wrong — verbal confidence is not an accuracy signal.

For any specific statistic AI provides ("73% of people..."), first ask for the precise source, then verify it directly before use.

After completing AI-assisted work (report, presentation, plan), ask AI itself to attack the work as a strict external critic before submission.

Before deploying AI-generated code, apply a specific security checklist (SQL injection, exposed secrets, input validation) — don't assume it's secure just because "it works".

After hearing your partner, reflect back three sentences: content, feeling, and underlying value.

Do not say "I know how you feel". Name the feeling with a precise word and verify it.

Paraphrase what you heard before responding — speaker confirms you heard.

After guessing what a close person feels, verify explicitly instead of trusting your hunch alone.

Write a prompt that gives the AI model precise context, role, task, and expected format.

Assign the model a specific background and expertise so its answer matches a real expert level.

Show the model 2-5 examples of the desired output style so it mirrors the pattern with high precision.

Ask the model to decompose the problem and show its reasoning before answering, raising accuracy on complex tasks.

Ask the model to write a numbered plan first and wait for your approval before executing.

State explicitly what you do not want in the output (words, styles, assumptions) to cut later editing.

Request output as JSON with a defined schema to enable automated processing without manual parsing.

After the model answers, ask it to critique its own answer and improve it before final delivery.

Ask the model to explain a complex idea to a 5-year-old, then raise the level gradually.

Learn 4 signals that reveal the model invented info instead of citing documented fact.

Do not use AI-sourced info before verifying it in two independent external sources.

The model does not know what happened after its training cutoff. Know each model's cutoff.

RAG (Retrieval-Augmented Generation) fetches real docs for the model before answering. Know the difference.

Before believing a "shocking" image, drag it to Google Lens or TinEye to find its true source and date.

Do not share any "shocking" news for 24 hours — breaking news is often corrected on day 1.

Use an LLM to draft an email in 30 seconds, then edit by hand for personal tone.

Convert a meeting transcript (60+ min) into decisions and action items in 2 minutes.

Turn a 200-page PDF into a Q&A conversation with exact-quote answers.

Don't pick the first title that comes to mind. Generate 10 with varied angles, then choose the best.

Ask the model to question you instead of answering, surfacing and deepening your understanding.

When an LLM helps you with content you publish or send, disclose it clearly.

Add "think step-by-step inside <thinking> tags before answering" — boosts complex-problem accuracy 30-40%.

In few-shot, deliberate variety (positive/negative/edge case) matters more than count.

Have the model play expert + student who asks "dumb" questions — extracts tacit knowledge.

"Max 80 words, don't use 'innovative', start with imperative" beats "make it short and clear".

Instead of rewriting the prompt, say "this part good. this part fix to X". 4 rounds = perfect output.

After each output, ask: "What are 3 weaknesses in this answer? How to improve?" — use its intelligence against itself.

Don't repeat instructions every message — set them once in System Prompt, save 70% writing time.

Don't know how to ask? Ask: "What's the best prompt to get X?" — it writes a prompt better than any expert.

Before accepting any AI answer, ask: source? dates correct? names real?

Every AI claim: search 3 independent sources. If absent from any, treat as hallucinated.

Before relying on AI in specialist field, show to human expert — 30% of time uncovers core error.

Any question on recent events/prices/releases: check model training date — ignoring = confident wrong answer.

"According to a 2019 study", "As Academy X stated", "approximate but accurate" — these are immediate-verify flags.

Break a big task into 4-6 sequential steps, each a separate prompt + human verification.

Use AI to expand research, summarize, surface angles — then YOU write the final text in your voice.

Explain your problem to the model out loud. You often find the solution while explaining, before its reply.

Feed one model's output to another and request harsh critique. Disagreements expose weak points.

Never write the same prompt twice. Save every successful prompt in a tagged, searchable file.

Set "system instructions" once in your account = every chat starts from your context automatically.

Before asking AI, paste relevant docs/refs. It answers from "your knowledge", not its imagination.

Every factual AI claim = ask for a source. No source = don't use.

Before publishing/sending: 5 mandatory verification questions. No exceptions.

Define 5 contexts where you'll NEVER use AI. Write them. Stick to them.

Use AI for a bad-fast initial draft, then YOU rewrite completely. Blank page is the killer.

After YOU write, ask AI for 3 improvements only. Don't let it "rephrase".

Identify 5 phrases/words that are "you". Ensure 2+ appear in any AI-assisted text.

"AI text didn't copy anyone" is myth. Ensure you don't publish someone's idea/sentence unattributed.

Every AI claim needs a live source. Never cite a source you haven't read.

Google your name every 3 months. First 10 results = your identity to the world.