Free to read · no account needed

18 min · foundations

What ChatGPT, Claude, and Grok can (and cannot) do

Goal: Know when to trust the answer, when to double-check, and when to stop using the chat.

Best for: Your first week with any AI chat. No tech background needed.

LLMs predict useful next words. They are not a truth database. Learn the safe jobs, the risky jobs, and the 60-second check before you trust an answer.

Why beginners skip this

Overrating live truth and actions while underrating drafts burns trust. Keep a can/cannot card and hand off what needs tools or humans. Mid-week cost shows up as rework, not as a missing bookmark.

Step 1 — Plain words: what is an LLM?

LLM means Large Language Model — the engine behind ChatGPT, Claude, and Grok. Think of it as a very strong autocomplete: it predicts useful next sentences from patterns in text it was trained on. It does not “look up a secret truth database” unless you turn on web search or attach a file.

  • Good at: drafting, rewriting, outlining, explaining, brainstorming, summarizing text you give it.
  • Risky at: fresh news, exact quotes, prices, medical/legal/finance decisions, anything you would bet money or reputation on.
  • Bad at: knowing your private calendar or inbox unless you connect those tools on purpose.

Step 2 — A 60-second test before you trust an answer

After any important answer, paste this follow-up (copy as-is, then fill the brackets):

Mark each claim in your last answer as:
A) Supported by text I already gave you
B) General knowledge that could be outdated
C) Guess / unsure
List claims under A / B / C. For every B or C, tell me what I should open to verify.
Example

Filled prompt

Mark each claim in your last answer as A / B / C.
Last topic: “Does Linear replace Jira for a 12-person SaaS eng team?”
For every B or C, tell me what page to open to verify.

What good output looks like

  • You see a short A/B/C list, not another essay
  • B/C items come with “open this docs page / ask your admin”
  • If almost everything is C, you treat the answer as a draft only
  • Pass/fail dated (Asia/Shanghai) sits next to the artifact

Step 3 — What to click in the product (no jargon trap)

If you need…Look for a button/toggle like…Why
Today’s news or pricesSearch / Browse / Web / ResearchWithout it, the model may invent outdated facts
Answer from your PDFAttach / Upload / + fileThen say “only use the file”
Safer long projectProjects (ChatGPT or Claude)Keeps files + standing instructions in one place

If you cannot find Search or Upload, check your plan’s help page — free plans sometimes hide these. You can still paste short text into the chat.

Step 4 — Practice: one safe task, one unsafe task

Do this step with a real artifact open—not from memory.

Task A (safe): Rewrite this paragraph for a busy manager in 5 bullets. Keep my meaning.
[paste paragraph]

Task B (unsafe alone): “What is our company’s exact refund policy?”
→ Instead attach the policy PDF, or say you will open the wiki yourself.
Example

Filled prompt

Rewrite for a busy manager in 5 bullets. Keep my meaning.
Paragraph: We will migrate checkout to the new payments API in November. Marketing needs a one-week freeze on promo codes during cutover. Support must use the new refund form after Nov 15.

What good output looks like

  • Five short bullets, no new promises invented
  • Dates and team names preserved
  • You would feel fine pasting this into Slack after a glance

Cannot = still useful

Knowing limits makes better asks—not cynicism.

Example

Filled prompt

Asks: (1) brainstorm subject lines (2) today's USD/HKD mid-market (3) medical dosage.
Label trust / check / stop. For check: what primary. For stop: refuse line.

What good output looks like

  • Labels trust/check/stop correctly for “Know when to trust the answer, when to double-chec”
  • Does not treat the model as a database or authority
  • Context pack stays ≤ one screen when asked
  • Pass/fail: you can explain the limit in one line

Red flag

Asking for live inventory with no tool; “just send the email” with no connector—correct the ask.

Save this card

Card: Trust drafts; check facts; stop on high-stakes and secrets—check it once this week, or it stays a bookmark. Keep this beside the tool; read before you start.

1. What I shipped: ___ 2. What I refused: ___ 3. What I will never-do again: ___ (date: Asia/Shanghai)

  • [ ] Tools/writes paused
  • [ ] Refuse-line dated
  • [ ] Artifact path saved
  • [ ] Next link chosen (or stop)

Before you start

Sit with three real asks from this week. You will label trust/check/stop for “What ChatGPT, Claude, and Grok can”—bring the asks, not slogans. Guide: foundations-what-an-llm-can-and-cannot-do. If you only have a blank practice box, stop and fetch the real job first.

  • Goal in one line: ___
  • Allowed tools: ___
  • Forbidden: ___
  • Human approve needed for: ___

How to use this week

This week: sort three real asks into trust / check / stop—drafts can feel right, facts need verify, high-risk or secrets stop cold. Ship the artifact out of chat (doc/PR/note) so the week leaves a file, not only a thread—date it Asia/Shanghai.

Calendar block (foundations-what-an-llm-can-and-cannot-do, optional): 25 minutes once this week—open the guide, run Practice once, date the refuse-line, then stop. No binge-reading counts as practice.

Pitfalls

PitfallFix
Confident tone = trueAsk for A/B/C labels
Asking “what is our policy?” with no fileUpload the policy or open it yourself
Using AI for medical/legal final callsUse it to prepare questions for a professional
Example

Filled prompt

Job: Know when to trust the answer, when to double-chec
Prior fail: [what missed the check]
Repair: keep the same Done; add one constraint + one concrete example from my notes.
Boundaries: draft only; label Unknowns [?]; no silent send.
Return: fixed output + the one-minute pass/fail I will mark.

What good output looks like

  • Names the prior fail without inventing a new job
  • Adds one constraint tied to “Know when to trust the answer, when to double-chec”
  • Still draft-only / Unknowns labeled
  • Pass/fail: you mark green on the same checklist (foundations-what-an-llm-can-and-cannot-do #4)

Practice now

Live card: one can / one cannot with a work example—not a meme (Asia/Shanghai).

Today: rewrite one overreaching ask into something the model can do, and mark what stays human. Date the rewrite (Asia/Shanghai).

Today card (copy) — foundations-what-an-llm-can-and-cannot-do: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

  • [ ] Artifact path written
  • [ ] Refuse example would say "yes"
  • [ ] Dated Asia/Shanghai

One sentence takeaway

Trust drafts; check facts; stop on high-stakes and secrets—check it once this week, or it stays a bookmark

  • You can point to a dated refuse-line
  • You have one artifact path from tonight
  • You can teach the red flag in one sentence

Next

Context & memory · Choose a model · Beginner mistakes · When AI sounds sure · Free vs paid

Pair with today: open the first Next link (foundations-context-windows-and-memory) only after you dated a refuse-line for this guide (Asia/Shanghai).

← Back to home · Learn · Tools · Compare · Start