AI assistants are useful collaborators, not oracles. This sheet helps you get value quickly while staying in control: treat every answer as a draft to verify, start with low-stakes tasks, and watch your usage so there are no surprises. Most of these habits transfer to any AI tool, not just Claude.
1. Start Here — Safe First Steps
Pick low-stakes tasks where you can judge the answer yourself.
- Claude.ai (web chat) is the gentlest start: ask questions, paste text, no install.
- Claude Code (CLI) is for editing real files/code — more power, so practice in a throwaway folder first.
- Begin with tasks you can check at a glance: summarize a paper you know, draft an email, explain code you wrote.
- Never paste secrets, credentials, or sensitive/unpublished data until you understand the tool's data policy.
Try this first
Summarize this abstract in 3 plain-language bullets,
then list any claims I should double-check before citing it.
[paste abstract]Copyable — replace the bracketed text.
TipYou stay in charge. The AI proposes; you decide what to keep. Nothing happens to your files unless you approve it.
2. Verify Everything (Trust, but Check)
Treat output as a confident first draft from a smart-but-fallible intern.
- AI can hallucinate: invented citations, fake function names, plausible-but-wrong numbers.
- Ask for sources and reasoning, then check the sources actually exist and say what's claimed.
- For facts: cross-check against a primary source. For code: run it and read it.
- Ask the same question two ways; inconsistent answers signal low confidence.
Verification prompts
For each claim above, label it [verified / uncertain /
needs a source] and give the exact reference. Do not
invent citations — say "I'm not sure" if unsure.What are the top 3 ways this answer could be wrong,
and how would I test each one?CautionA real DOI or URL can still be fabricated. Click through and confirm before you cite or ship anything.
3. Cost & Token Awareness
"Tokens" are chunks of text (~¾ of a word). On a subscription (Pro/Max) you spend a usage budget against per-session and weekly limits; only API users see per-token billing.
- Plans have per-session and weekly limits; bigger models cost more per token (a "premium").
- Long pasted context + long replies = fast burn. Trim inputs; ask for concise output.
- Start a fresh chat for a new topic — old history is re-sent each turn and adds up.
- Other tools (OpenAI Codex, Google Antigravity) have their own limits/pricing — compare before committing.
Cost-control prompts
Answer in under 150 words. If you need more, ask first.Give me just the changed lines, not the whole file.TipUse a smaller/faster model for simple tasks and save the big model for hard reasoning — see the next card.
4. Model & Effort Choice
Match the engine to the job. Don't pay for Opus to fix a typo.
| Choice | Use when |
|---|---|
| Haiku | Quick, cheap: short Q&A, formatting, simple edits. |
| Sonnet | Everyday default: solid balance of speed, cost, quality. |
| Opus | Hard reasoning, tricky debugging, dense synthesis. |
| Low effort | Fast factual or lookup tasks. |
| High / xhigh | Multi-step problems worth slower, deeper thinking. |
TipDefault to Sonnet at medium effort. Escalate to Opus/high only when the answer matters and you're not satisfied.
5. Ask for Artifacts & Diagrams
For longer or denser answers, request a visual or a self-contained file.
- Ask for an HTML artifact (a table, checklist, mini-report) you can save and reuse offline.
- Ask for an SVG diagram to see a workflow or concept instead of a wall of text.
- Great for skeptics: a diagram makes the AI's reasoning inspectable.
Artifact prompts
Turn this comparison into a clean HTML table I can
save as a single offline file.Draw an SVG flowchart of this process, with a short
caption explaining each step.TipIf a reply is getting long, say "make this an artifact" — it's easier to scan, save, and check.
6. Privacy, Security & Trust
Safe habits matter even at the newcomer stage.
- Assume anything you paste leaves your machine. Keep secrets, PII, and embargoed data out.
- Read the Terms of Use for data retention and whether your inputs train future models.
- MCP servers (connectors to your files, web, databases) are powerful — only add ones you trust, grant least access, and review what they can do.
- In Claude Code, read the diff and approve each action rather than auto-accepting changes.
CautionA connector or agent that can edit files or browse the web can also be misused. Start with read-only, narrow scope, and watch what it touches.
7. Do / Avoid
Do
- Start with tasks you can verify yourself.
- Ask for sources, reasoning, and confidence levels.
- Run and read any code before trusting it.
- Keep prompts and outputs concise to save tokens.
- Use a fresh chat per new topic.
- Match model/effort to task difficulty.
- Approve each file change deliberately.
Avoid
- Pasting secrets, credentials, or sensitive data.
- Citing AI-provided references without checking them.
- Auto-accepting edits you haven't read.
- Using Opus + max effort for trivial tasks.
- Carrying one giant chat across unrelated topics.
- Installing MCP connectors you don't recognize.
- Treating a confident tone as proof of correctness.
8. Quick Reference — Web Chat vs CLI, and a Skeptic's Workflow
| Question | claude.ai (web chat) | Claude Code (CLI) |
|---|---|---|
| Best for | Research questions, drafting, learning, artifacts | Editing real files & codebases, automation |
| Risk level | Low — it can't touch your files | Higher — it can change files (you approve each step) |
| Newcomer move | Start here | Try later, in a throwaway folder |
A skeptic's 4-step loop
- 1. Scope small — one clear, checkable task.
- 2. Ask + request reasoning — "show your work and your sources."
- 3. Verify independently — run the code, click the link, check a primary source.
- 4. Keep or discard — accept only what survives the check; iterate on the rest.
TipThese habits are model-agnostic — they work the same with other AI coding and chat tools, so your skill transfers.