Think Mode (Extended Thinking)

Module 6.1: Think Mode (Extended Thinking)

Estimated time: ~30 minutes

Prerequisite: Module 5.3 (Image & Visual Context)

Outcome: After this module you will know extended thinking is already on, how to raise or lower it with effort levels, when to reach for ultrathink, and how to read the reasoning Claude Code hides.

1. WHY — Why This Matters

Half the Claude Code advice online tells you to type “think step by step” or “think harder” to unlock a deeper mode. Those phrases do nothing: Claude Code “passes other phrases such as ‘think’, ‘think hard’, and ‘think more’ through as ordinary prompt text and doesn’t recognize them as keywords.”

So you paid for magic words while the real controls — the effort level and the ultrathink keyword — sat unused, and thinking was on all along. This module replaces the folklore with four switches that exist.

2. CONCEPT — Core Ideas

Extended thinking is on by default. There is no mode to enter. You control how much reasoning Claude spends (effort level, session-wide), whether one turn gets extra (ultrathink), whether it happens at all (Option+T), and whether you see it (Ctrl+O).

The effort level is the main dial. This is the old think-vs-act matrix, re-axed onto levels that exist:

The task is…	Effort
”short, scoped, latency-sensitive… not intelligence-sensitive”	`low`
cost-sensitive, “can trade off some intelligence”	`medium`
everyday coding: “balances token usage and intelligence”	`high`, the default
worth “deeper reasoning at higher token spend”	`xhigh`
demanding, but “prone to overthinking” — “test before adopting broadly”	`max`
big enough to orchestrate: “`xhigh` … plus dynamic workflows”	`ultracode`

high is “the default on every model except Opus 4.7”, which defaults to xhigh.

ultrathink is the one keyword Claude Code recognizes, and it lasts a single turn: “Include ultrathink anywhere in your prompt to request deeper reasoning on that turn without changing your session effort setting… The effort level sent to the API is unchanged.”

graph TD
    A[Your prompt] --> B{Thinking on?<br/>Option+T · alwaysThinkingEnabled}
    B -->|off| E[Plain response]
    B -->|on, the default| C[Effort level<br/>low · medium · high · xhigh · max]
    A -->|contains ultrathink| D[+ in-context instruction<br/>this turn only]
    C --> F[Reasoning, collapsed]
    D --> F
    F -->|Ctrl+O| G[Reasoning visible]

3. DEMO — Step by Step

All steps run in a lab repo holding src/math.js and one test file.

Step 1: Read the effort level you are on

Start claude and look above the prompt box.

# Output may vary
 ▐▛███▛█   Claude Code v2.1.278
▝▜██████▀  Opus 5 (1M context) · Claude Max
  ▝▝ ▝▝    ~/cc-lab
                                                                                  ● high · /effort
────────────────────────────────────────────────────────────────────────────────────
❯

● high · /effort is your current level. You never turned thinking on — it already was on.

Step 2: Open the effort slider with /effort

Type /effort and press Enter.

# Output may vary
   Effort
                   Faster                                                 Smarter
                   ────────────────────▲──────────────────────┆──────────────────
                   low     medium     high     xhigh      max       ultracode
                                                                xhigh + workflows
   ←/→ to adjust · Enter to confirm · s for this session only · Esc to cancel

The slider shows the levels this model has — Opus 4.6 and Sonnet 4.6 have no xhigh. /effort <level> sets one; /effort auto clears the saved level for that model.

Step 3: The on/off switch, Option+T

Press Option+T (macOS) or Alt+T (Windows/Linux).

# Output may vary
  Toggle thinking mode
  Enable or disable thinking for this session.
  ❯ 1. Enabled ✔  Claude will think before responding
    2. Disabled   Claude will respond without extended thinking
  Enter to confirm · Esc to cancel

Enabled is already checked. /config saves it as alwaysThinkingEnabled in ~/.claude/settings.json.

Step 4: The one keyword that works

ultrathink: does the guard in divide() in src/math.js also reject -0? Answer in one sentence.

# Output may vary
❯ ultrathink: does the guard in divide() in src/math.js also reject -0? …
⏺ I'll read the file.
  Thought for 2s, read 1 file
⏺ Yes — b === 0 is true for -0 as well (strict equality doesn't distinguish the two zeros,
  unlike Object.is), so divide(1, -0) also throws instead of returning -Infinity.

Thought for 2s is all you get: “Claude Code collapses thinking output by default.”

Step 5: Expand the reasoning, Ctrl+O

Press Ctrl+O to toggle the transcript viewer.

# Output may vary
                                                                  10:00 PM claude-opus-5
⏺ I'll read the file.
⏺ Read(/Users/luatnq/cc-lab/src/math.js)
  ⎿  Read 6 lines
∴ Since -0 equals 0 in the comparison, it correctly rejects -0 too.
────────────────────────────────────────────────────────────────────────────────────
  Showing detailed transcript · ctrl+o to toggle · ? for shortcuts          verbose

The ∴ line is the thinking summary. This capture used { "showThinkingSummaries": true } in .claude/settings.json, because “interactive sessions on the Anthropic API receive redacted thinking blocks by default.”

Step 6: Measure the difference, headlessly

# docs: cli-reference
claude --effort low  -p "List the edge cases divide() in src/math.js does not handle." \
  --output-format json --permission-mode default --allowedTools "Read"
claude --effort high -p "List the edge cases divide() in src/math.js does not handle." \
  --output-format json --permission-mode default --allowedTools "Read"

Both runs are read-only, so --allowedTools "Read" is their whole permission budget. The two JSON objects, same prompt and repo: result 1,399 characters at low vs 2,494 at high; num_turns 2 vs 4; duration_ms 10,615 vs 40,056. Output may vary — one pair of runs on one machine. Compare length, turns and content, never a thinking-token count: Claude Code does not print one.

4. PRACTICE — Try It Yourself

Exercise 1: Prove the ladder is dead

Goal: Confirm “think step by step” is ordinary prompt text and ultrathink is not.

Instructions:

In a repo of yours, run claude and ask a question prefixed with think step by step:.
Press Ctrl+O and note the reasoning.
Ask the same question prefixed with ultrathink:. Press Ctrl+O again.
Compare the two transcripts.

Expected result: Both turns think, because thinking is on by default. Only the ultrathink turn got an extra instruction.

💡 Hint

Do not look for a “mode” indicator. Compare reasoning length in the transcript viewer — step 1’s phrase reached the model as plain text.

✅ Solution

The first prompt is like any other — Claude Code “passes other phrases such as ‘think’, ‘think hard’, and ‘think more’ through as ordinary prompt text.” The second is recognized: it “adds an in-context instruction. The effort level sent to the API is unchanged.” ultrathink buys a deeper turn; the folk phrases buy only tokens.

Exercise 2: Find your own effort floor

Goal: Pick the cheapest effort level that still solves a task you run.

Instructions:

Choose a real, repeatable task (for example “summarize what changed in the last commit”).
Run it with --effort low, --effort medium and --effort high, each with --output-format json, saving every result.
Compare the result fields for correctness, not length.
Pin the winner with /effort <level> or the effortLevel setting.

Expected result: A note saying which level you will use for this task class.

💡 Hint

low is for work “not intelligence-sensitive”. If low gets the right answer, high is wasted spend.

✅ Solution

For mechanical tasks low often ties high, so pin it per model with modelSettings. The docs are specific: “set effortLevel to low, medium, high, or xhigh as the default for models without one. max isn’t accepted as a level in either key, and ultracode has its own ultracode key.” So max must come from /effort or --effort.

5. CHEAT SHEET

Control	Where	What it does
`/effort` · `/effort auto`	Session	Slider (`s` = session only) · clear the saved level
`/effort <level>`, `--effort <level>`	Session, CLI	`low`, `medium`, `high`, `xhigh`, `max`, `ultracode`
`effortLevel`, `modelSettings`	settings.json	Default, and per model: `low`–`xhigh` only
`ultracode`	settings.json	Its own key — `xhigh` plus dynamic workflows
`ultrathink`	Prompt	Deeper reasoning, that turn only
`Option+T` / `Alt+T`	Session	Toggle extended thinking
`alwaysThinkingEnabled`	settings.json	Thinking off for every session
`MAX_THINKING_TOKENS=0`	Env var	Thinking off on the Anthropic API
`Ctrl+O`	Session	Toggle the transcript viewer
`showThinkingSummaries`	settings.json	Full summaries instead of redacted blocks

Defaults: thinking on; effort high except Opus 4.7 (xhigh). Thinking cannot be turned off on Fable.

6. PITFALLS — Common Mistakes

❌ Mistake	✅ Correct Approach
”think step by step” to unlock a deeper mode	Thinking is already on. Raise `/effort`, or use `ultrathink`
Quoting a “thinking tokens: 2,768” figure	Claude Code prints no such count. Compare length and turns
Assuming `ultrathink` raises your effort level	”The effort level sent to the API is unchanged”
Running everything at `max`	`max` “may show diminishing returns and is prone to overthinking”
`"effortLevel": "max"` in settings.json	”`max` isn’t accepted as a level in either key”. Use `/effort max` or `--effort max`
Expecting `Option+T` to work on Fable	”You can’t turn thinking off on Fable models”
Assuming collapsed reasoning is free	”You are charged for all thinking tokens generated, even when collapsed”

7. REAL CASE — Production Story

Scenario: A Vietnam-based fintech team building payment reconciliation: VND amounts (no decimals), three banks with different API response formats (VietcomBank, Techcombank, MBBank), and State Bank of Vietnam (SBV) audit requirements.

Problem: They asked Claude Code to “implement payment reconciliation logic” at the default effort, in a session already holding two unrelated tasks. The result assumed USD-style minor units (amount * 100), one uniform bank response shape, and skipped the SBV state machine and audit trail. Five rounds of corrections followed, piling up in the same polluted context.

Solution: The team lead ran /clear, raised the session with /effort xhigh, and rewrote the prompt to carry the constraints the first attempt had to guess: VND is an integer currency, each bank has its own field names, SBV requires a state machine plus an audit trail, settlement windows differ per bank. The hardest sub-problem — mismatched settlement windows — got its own turn prefixed with ultrathink.

Result: Integer VND handling, a per-bank adapter layer, and an explicit state machine with audit events. The rule they wrote into CLAUDE.md was not “add think to your prompts” — it was: clear the context, put the constraints in the prompt, reach for effort and ultrathink deliberately.

Next: Module 6.2: Plan Mode →