Fast is for motion. Think is for load-bearing answers. Mixing them up wastes time or ships a guess.
Why this matters
On grok.com you can choose how hard the model works before it speaks. The control sits near the composer. The label has been Fast vs Think. Names and extra modes will move — Multi-agent and other heavy options show up when xAI ships them. The decision does not move.
If you Think a rewrite of a two-line text, you wait for no gain. If you Fast a decision with money, safety, or reputation on it, you get fluency instead of scrutiny. Fluency feels like thought. It is not.
This lesson is one split you can defend: which jobs get speed, which jobs get a slower pass, and when you do both.
Teach
Fast is the default for drafts and classification. Use it when you already know what “good” looks like and you need volume or a first cut. Emails. Rewrites. Checklists from facts you already trust. “Turn these eight lines into a table.” “Give me three subject lines.” You will read the output in seconds and you can see if it is wrong.
Think is for work that has to survive a question. Use it when the job has tradeoffs, hidden steps, or a chance to be confidently wrong. “What breaks if we move the pour to 13:00.” “Compare these two options and name what would change my mind.” “Find the hole in this plan.” Think spends time (and, on some plans, more of your quota) to reason before it answers. That only pays if you will use the reasoning.
A third pattern, usually the right one for real work: Fast, then Think. Fast builds the artifact. Think audits it. Same thread.
Stay in Think. Do not rewrite the draft unless a line is false.
Audit the draft above for:
- invented facts
- missing constraints from my brief
- the single biggest operational risk
Return a 6-line audit, then a pass/fail.
That second pass is cheap insurance. It is not a personality. Do not ask Think to “make it nicer.”
How to choose, in practice:
- You will edit in under two minutes → Fast.
- Someone else will act on this without you in the room → Think, or Fast-then-Think.
- The facts might be stale (prices, laws, weather, a public statement) → Search, which is the next lesson. Thinking harder will not refresh the world.
- The job is a picture → Imagine, not Think. Reasoning will not draw the layout.
- You are stuck in a loop of taste (“more punchy”) → Fast, and then stop. Think will not find your taste for you.
Do not worship the slow mode. A long think on a bad brief is a long think on a bad brief. ICC still comes first. If you cannot name the artifact, Fast and Think will both wander.
Plans and rate limits change. If Think is gated or exhausted, say so in your notes and do the audit yourself with a Fast prompt: “List facts you are not sure about.” That is a weaker substitute. It is still better than pretending.
Ignore chrome. If the composer shows Auto, Fast, Think, or something heavier, map it back to the same question: does this job need speed, scrutiny, or both?
Worked example
A coordinator needs a site-visit plan: three people, one truck, a two-hour window, a locked gate.
Fast brief (same ICC you would use anyway):
Intent: A timed plan I can text the crew tonight.
Context: Tuesday 08:00–10:00, north gate, one truck, three people. Gate code with dispatch. We must leave a lane for the county inspector at 09:00. Tools: two GPS rovers, spare battery, printed control sheet.
Constraints: Under 15 lines. Times in local clock. No pep talk.
Write the plan.
Fast will give you a clean list. It may also skip the collision: three people, one truck, inspector lane, gate code. It may invent a second vehicle because “that would help.”
Think pass, same thread:
Think through this as a field lead. Do not add staff I do not have.
Questions to answer first, then revise the plan:
- Who is out of the truck when the inspector needs the lane?
- What happens if the gate code fails?
- Which task gets cut if we start 20 minutes late?
Mark any line in the Fast plan that assumed a resource I did not give.
The Think reply should change the plan or explicitly say the Fast plan already held. Either result is useful. A Think reply that only restates the Fast plan in longer sentences did not earn the wait.
Your turn
Open grok.com. One new thread. You will send the same ICC brief twice: once in Fast, once in Think. Then you will ask Grok to compare them. You cannot pass this by reading about modes.
If your plan has no Think control, say so in the artifact and still send the second message: “Answer again, slower, and show the risks you skipped.” The comparison is the work.
Exercise a3-l01-e1
| Field | Value |
|---|
| whereToDo | grok.com |
| time | 16 min |
Copy-ready prompts
Message 1 — set the composer to Fast. Fill the brief with a real decision from your week that has at least one tradeoff (time, money, people, or risk). Not a poem. Not “write a joke.”
Mode check: I believe I am on Fast. Confirm in one word at the top.
Intent: [a plan, comparison, or recommendation a person will act on]
Context: [facts, including the resource limits]
Constraints:
- Under 20 lines
- Do not invent people, money, or equipment
- Name the option you recommend
Write the artifact.
Message 2 — same thread. Set the composer to Think.
Mode check: I believe I am on Think. Confirm in one word at the top.
Same brief as my first message. Do not read my taste. Re-solve it.
Show a short reasoning pass, then the artifact.
Flag any Fast-line that looks fluent but unsupported.
If the UI will not label the mode in the reply, you still switch the control. The next message asks Grok to work from what it can see.
Message 3 — either mode is fine:
Write a Mode Split with exactly these lines:
- Fast recommendation:
- Think recommendation:
- Where they agree:
- Where they disagree:
- One Fast fluency risk (a line that sounded sure):
- When I should have used Fast only:
- When I should have used Think (or Fast-then-Think):
- Session stamp: today's date and "Mode split written by Grok in this thread."
Do not invent a disagreement. If they match, say they match.
Artifact to paste
- The Mode Split (all lines, including the session stamp).
- The first line of each of the two artifacts (Fast and Think), so it is clear both ran.
- Your rule, one sentence: “I will use Think when ___ ; I will stay on Fast when ___.”
Rubric
- Pass: Two mode checks or two clearly different passes exist. The Mode Split is Grok-authored and stamped. Your rule is specific to a kind of job, not “I will use both equally.” The underlying job has a tradeoff (not “summarize this sentence”).
- Fail: One mode only. A split you wrote from this lesson without sending the prompts. A job with no decision (a haiku, a compliment). You report “they felt the same” and paste nothing.
Common failure
The common failure is Think as a superstition. Every message goes through the slow door because it feels serious. You will burn quota, wait, and still send a sloppy brief. Serious is the brief and the read, not the spinner.
The twin failure is Fast for cover. You need a recommendation you can blame on the model. Fast will give you one. It will not have checked the thing that hurts. If a person will drive, climb, spend, or sign, you owe them a Think pass or your own checklist. Prefer your own checklist if you do not understand the Think reply.
What’s next
You can choose a speed. Next: the other desk tools — Search when the world moved, Imagine when a picture is the artifact, files when you have source material, and a thread that holds the job together.
