AI bot guides: hallucinations, credits, choosing a bot, files, prompts, privacy

1. Know where it fails

Exact numbers, prices, dates, quotes, citations, legal and medical specifics, news after its training cutoff, and claims about your accounts or files it hasn't opened.

2. Make it show the source

Bad: "What does Claude Pro cost?"

Good: "What does Claude Pro cost? Quote the line from claude.com/pricing and give the URL. If you can't load it, say so."

Why: you now have something checkable, and permission to say "I don't know".

3. Open the link

Ctrl+F the quoted sentence. Fake or wrong citations are common: the link exists but doesn't say that.

4. Separate checked from assumed

Ask: "Mark each claim CHECKED (with source) or ASSUMED."

5. Demand proof for actions

"Done" is not proof. Ask for the post URL, the deploy ID, the sent-email link, or a screenshot, and open it.

6. Cross-check what matters

For money, health, law or anything public: ask a second bot or a primary source. Disagreement = dig in.

1. Find what's using it

Open your usage page (Claude: Settings > Usage; Cursor/Grok Bot: usage page; Copilot: billing). Look for scheduled jobs and long agent runs first.

2. Kill unattended jobs

Scheduled tasks/routines run whether or not anything changed. Ask: "List every scheduled job, how often it runs, and what it last did." Turn off any you don't read.

3. Start fresh chats

Each reply re-processes the conversation. A 200-message chat costs more per question than a new one. Paste a 5-line summary into a new chat instead.

4. Right-size the model

Quick rewrites and lookups don't need the biggest model or deep research. Save those for hard problems.

5. Give the whole task at once

Bad: 10 messages of "now make it shorter… now add a title…"

Good: one message with length, title, tone and format.

6. Stop loops early

If an agent fails the same step twice, stop it and change the instructions. Failed runs cost the same and are not refunded.

Writing and editing

ChatGPT, Claude, Gemini. Try the same prompt in two and keep the one you edit least.

Research with sources

Perplexity or any bot with web search on. Always open the sources.

Your Google files and Gmail

Gemini on a Google AI plan works inside Gmail and Docs.

Microsoft 365 documents

Microsoft Copilot, inside the Office apps your plan includes.

Code in a real project

Cursor, GitHub Copilot, Claude Code, Codex: they read the repo and run tests.

Doing things for you

An agent with connectors and approvals, like Grok Bot. Approvals on for posting, email, buying, deleting.

Images

Midjourney, ChatGPT or Gemini image tools. Check each one's commercial-use terms.

1. Decide what's off limits

Passwords, card numbers, one-time codes, SSNs, medical records of others, client data under NDA.

2. Redact first

Replace names and numbers with placeholders (CLIENT_A, ACCT_1) before upload.

3. Check training settings

ChatGPT: Settings > Data controls. Claude: Settings > Privacy. Gemini: Gemini Apps Activity. Turn off training if you don't want it.

4. Use official connectors

Connect Drive/Gmail through the bot's connector, never by sharing a password. Grant the smallest access offered.

5. Ask precise questions

"List every date and dollar amount in this PDF with the page number." Page numbers let you verify.

6. Clean up

Delete chats/files you don't need, and revoke connectors you stopped using.

From each bot's page on this site. This isn't a ranking. Try the free tier first.

The 4 parts

1. Task: what to do
2. Context: who it's for and why
3. Format: the length, a list or table, the tone
4. Check: "list your sources" or "say if you're unsure"

Before and after

Weak: "write about flood safety"

Better: "Write 5 bullet points on flash flood safety for drivers, plain English, and cite weather.gov."

• Never paste passwords, card numbers or one-time codes into a chat.
• Read each bot's data policy. Anthropic, for example, lists model training as opt-out on consumer plans.
• Connect only the accounts a task needs, and remove the ones you stop using.
• Make posting, emailing and buying need your approval.
• Use the official site or app only. Cursor, for example, warns that subscriptions from resellers may be fraudulent.

1. Ask for the link, the file or a screenshot. "Done" isn't proof.
2. Open the source and find the exact sentence.
3. Ask "when did you last check?" Status reports go stale.
4. Watch for confident numbers with no source.
5. Spot-check scheduled jobs yourself every week.

A real example: one Grok Bot user's emergency alert system was reported as working, but it had been dead for days. See the Grok Bot tutorial, Chapter 4.

How plans charge

• Rolling windows: Claude resets every 5 hours, with weekly caps on paid plans.
• Gemini refreshes every 5 hours, up to a weekly limit.
• Cursor and Grok share a usage pool that refreshes weekly.
• GitHub Copilot uses credits, where 1 credit = $0.01.

Controls

• Set a spending cap where the bot offers one.
• List and turn off scheduled jobs you didn't ask for.
• Remember that failed runs cost the same as good ones, and most plans don't refund them. Claude and Google say so on their pricing pages.