Extended Thinking

 Without thinking blocks carried back, Claude is re-reasoning from scratch each turn. On a complex plan it usually reaches the same conclusion — but not always. The thinking block carry-back ensures Claude is continuing the same reasoning thread, not starting a fresh one that happens to look similar.

Think of it like this:

Without carry-back:  Claude re-derives the plan each turn
                     → Usually same result, occasionally diverges

With carry-back:     Claude continues its original plan each turn
                     → Deterministically consistent

---

When does that difference actually matter?

Honestly, for simple 2-3 step tool loops — it probably doesn't. Your turns example would work fine.

It matters when:

• The plan has 6-8 dependent steps
• Claude made a conditional decision early on ("if A returns X, skip C")
• That condition is not visible in the assistant message, only in the thinking block
• By turn 5, Claude has forgotten its own conditional and executes C anyway

What will happen if we don't send the thinking block back in the Turn?

Without thinking block carry-back:

Turn 1: Claude reasons A→B→C→D→E→F  (full reasoning)
Turn 2: Claude reasons A→B→C→D→E→F  (full reasoning again)
Turn 3: Claude reasons A→B→C→D→E→F  (full reasoning again)
...repeats every turn

Wasteful, and risks diverging on complex plans.

---

With thinking block carry-back:

Turn 1: Claude reasons A→B→C→D→E→F  (full reasoning)
Turn 2: Claude continues from where Turn 1 left off  (picks up at B)
Turn 3: Claude continues from where Turn 2 left off  (picks up at C)
...progresses forward each turn

---

Two benefits this gives you:

Efficiency — Claude isn't re-deriving the same plan repeatedly. It already knows where it is in the plan.

Consistency — Any conditional decision Claude made in Turn 1 ("if A returns X, skip D") is preserved in the thinking block. Without carry-back, that condition might be re-derived differently in Turn 4.