Free to read · no account needed

16 min · claude

When to use extended thinking (and when it wastes time)

Goal: Turn on deeper reasoning for hard tradeoffs and multi-constraint problems—not for rewriting a three-line email.

Best for: Planning, debugging logic, comparing options with constraints. Product names for “thinking” modes vary.

Longer thinking helps on gnarly tradeoffs. It is slower and still fallible. For simple GFB tasks, normal mode wins.

Longer thinking can help on gnarly problems. It is slower and still fallible. For simple GFB tasks, normal mode is enough. Always keep Boundaries: no invented facts.

Extended thinking helps on multi-constraint puzzles—not on “rewrite this email nicer.” Spend the slower mode where mistakes are expensive.

Why beginners skip this

Extended thinking on greetings wastes money; hard tradeoffs with it off waste judgment. Match depth to how expensive a wrong answer is. Mid-week cost shows up as rework, not as a missing bookmark.

Step 1 — Good fits

  • Multi-constraint planning (“must hit A, B, cannot do C”)
  • Explaining a tricky bug after a minimal repro
  • Comparing 3 options with an explicit rubric

Poor fits: subject lines, tone polish, obvious summaries of text you pasted.

Step 2 — Prompt pattern

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

Use careful reasoning.
Goal: [ ]
Constraints: [list]
Return: decision + 5-bullet why + risks.
Do not invent data. Mark Unknown.
Example

Filled prompt

Tradeoff: ship dark mode October vs wait for design tokens.
Deeper reasoning ON. Options + risks + what would change my mind.
Reject single confident winner with no tradeoffs.

What good output looks like

  • Labels task A no-deep / task B yes-deep with a one-line why
  • Cites cost of mistakes (money, trust, irreversibility)
  • Five bullets you can paste into a personal rule card
  • Does not recommend deep mode for every chat
  • Keeps beginner wording—no vendor feature tour

Step 3 — Still verify

Deep thinking is not a source. Numbers and citations still need your tabs.

Time-box the deep pass

If thinking mode is available, give it a clear stop: “Return a decision in ≤12 bullets; then stop.” Endless rumination is not rigor.

Think carefully, then stop: decision + 5 why-bullets + 3 risks. No essay.
Example

Filled prompt

Write a prompt that forces the model to show a short plan before answering a trade-off question about pricing experiments—without inventing revenue numbers.

What good output looks like

  • Forces a short plan before the trade-off answer
  • No invented revenue, headcount, or market figures
  • Trade-off framed with at least two real options you named
  • States what would falsify the recommendation
  • Stays under your length/format boundaries

Save this card

Card: Deep mode for high cost-of-wrong tasks; fast mode for wording—match depth to stakes—one live pass/fail, or it was only a title. 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

Open the Claude Project you will keep. Confirm standing instructions + the knowledge file for “Turn on deeper reasoning for hard tradeoffs” are current. Guide: claude-extended-thinking-when-to-use. 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: run the same hard tradeoff once with extended thinking and once fast; keep depth only if the stakes warrant it. Note which you shipped on a live artifact—not a blank demo. Write one refuse-line and date the note (Asia/Shanghai). If you only skimmed claude-extended-thinking-when-to-use, put 25 minutes on the calendar and redo once.

Match stakes

RSVP email ≠ multi-constraint scheduling. Spend slow mode where mistakes are expensive.

Cost of wrong > cost of tokens

If a wrong answer means a bad hire, a bad price, or a public post you cannot pull, pay for a plan-first pass. If it is subject-line polish, stay fast. Do not confuse longer output with better judgment—still verify facts.

Example

Filled prompt

Everyday task: shorten email subject lines.
Should deeper thinking be ON? Argue no—and show when I would turn it on instead.

What good output looks like

  • Pulls from Project/file named in the ask for “Turn on deeper reasoning for hard tradeoffs and mu”
  • Standing instructions respected; no silent scope creep
  • Decisions quoted or labeled Unclear — never “sounds good”
  • One-minute check: does output match claude-extended-thinking-when-to-use Done?
Example

Filled prompt

Task: fix comma in email subject.
Should deep thinking be ON? Argue no; give a real tradeoff when yes.
Context: real workstream this week (Asia/Shanghai). Name actual files/people.
Constraints: no invented metrics/owners/prices; label Unknowns; stop before send/delete/pay.
Done check: I can mark pass/fail in one minute on the output.

What good output looks like

  • Standing instructions respected; no silent scope creep
  • Decisions quoted or labeled Unclear — never “sounds good”
  • One-minute check: does output match claude-extended-thinking-when-to-use Done?
  • Pulls from Project/file named in the ask for “Turn on deeper reasoning for hard tradeoffs and mu”

Red flag

Extended thinking on greetings; hard tradeoffs with thinking off; no record of when it helped—log one comparison.

Pitfalls

PitfallFix
Thinking mode for every chatReserve for hard tradeoffs
Trusting longer = truerVerify claims
Skipping constraintsList them
  • Extra pitfall: treating the bookmark as the practice—open a live artifact, run Practice now, and date a pass/fail (Asia/Shanghai).

Practice now

Live card: one Claude artifact/diff/email pass for “When to use extended thinking (and when it wastes time)” with Blind before share (Asia/Shanghai).

Today: write two prompts: (A) a tone rewrite that should stay fast; (B) a trade-off that must show a short plan before the answer. Run B with plan-first; reject invented revenue numbers.

Today card (copy) — claude-extended-thinking-when-to-use: 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

Deep mode for high cost-of-wrong tasks; fast mode for wording—match depth to stakes—one live pass/fail, or it was only a title

  • 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

Rubrics · Eval outputs · Plan-act-check · Claude Projects knowledge

Pair with today: open the first Next link (prompt-rubrics-and-acceptance-tests) only after you dated a refuse-line for this guide (Asia/Shanghai).

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