Adaptive Thinking, /effort, Output Styles, Plan Mode

Dial reasoning depth and tone separately

Claude 4.x replaced the deprecated Extended Thinking budget with Adaptive Thinking: the model decides how deeply to reason based on problem complexity, not on a token budget you set blind. You guide it with /effort low|medium|high|max — low for quick answers, max for architecture decisions on Opus.

Prompt keywords also bias depth: "think hard about…", "think step by step", "reason through all the tradeoffs". The strongest signal is the magic word "ultrathink" — explicitly asks for max thinking tokens. Use it for the gnarliest questions, not for everything.

Output styles change tone, not capability: Explanatory (default, explains reasoning), Concise (shorter, direct), Formal (no contractions, professional), Teaching (educational walk-through). Set in ~/.claude/settings.json or via /config.

Plan Mode (Shift+Tab) forces an implementation plan before any file is touched. Read the plan, redirect if needed, then approve. Indispensable for risky changes.

Code

Effort + magic word together·bash

# Mid-session
/effort high

# Then send the gnarly question
"ultrathink: design the cache invalidation strategy for our
 multi-region API given eventual consistency in Postgres
 read replicas and 60s TTL on Cloudflare edges"

# Or set effort at launch
claude --effort max --model opus \
  "design a distributed rate limiter for 100k req/s"

# Drop back when you're done
/effort medium

Plan Mode for safety·text

# In any session
press Shift+Tab            # toggles Plan Mode (banner appears)

"refactor the auth module to use JWT RS256 instead of HS256"

# Claude generates a plan:
# 1. Read src/auth/, identify HS256 references (4 files)
# 2. Generate RSA key pair via crypto.generateKeyPairSync
# 3. Update src/auth/sign.ts and src/auth/verify.ts
# 4. Add migration helpers for in-flight HS256 tokens
# 5. Update tests/auth/*.spec.ts
# 6. Document the rollout in docs/auth-migration.md
#
# [Approve]  [Redirect]  [Cancel]

# Nothing touched until you click Approve

Adaptive Thinking, /effort, Output Styles, Plan Mode

Dial reasoning depth and tone separately

Code

External links

Exercise

Progress

Comments 0