Chained claude -p prompts: same 30/30, 1.55× the cost

The test isolates one variable: the same six-line step list, byte-identical, sent either as one numbered prompt or as six sequential prompts that carry the conversation forward. Scoring was frozen before any run via hash values, and a reference implementation confirmed all six checks were simultaneously satisfiable. Across five trials per method, neither approach broke the other's prior achievements—the stepwise per-step counts were [2, 3, 3, 4, 5, 6] in all five trials, with zero regressions.

The cost gap traces to conversation re-reading rather than upfront prompt size. The first exchange carried only 1,567.2 more input tokens in the bundled version on average, but each later step in the split version re-read the accumulated history. The author notes this breakdown was only extracted for one trial's six calls, so the link between token-type mix and the 1.55× versus 2.51× multiplier mismatch remains unverified.

A separate finding concerns measurement hygiene: resumed claude -p calls report cumulative cost fields, and summing them produced a spurious 6.3× figure before correction to 1.55×. The author also observed that shell commands were denied in all 35 runs under acceptEdits—45 Bash denials total—so both methods reached 6/6 using file-writing tools alone, without self-verification.