What a "be terse" system prompt gets you, vs. Atelier's runtime.
Before anyone installs a tool, the free thing to try is telling Claude to write less. We tested that head-to-head against Atelier's actual runtime, on the same 20-prompt Telegraphic Q&A set, because it's the one "competitor" every reader can reproduce in thirty seconds with no install at all.
“Respond terse like smart caveman. All technical substance stay. Only fluff die.”
| Prompt | Baseline | Atelier (full) | Atelier (telegraphic only) | Caveman (persona only) |
|---|---|---|---|---|
| eval-02 Q&A prompt, no repo | 1258 | 1040 (-17%) | 997 (-21%) | 196 (-84%) |
| async-refactor Callback -> async/await refactor | 785 | 153 (-81%) | 192 (-76%) | 193 (-75%) |
| error-boundary React error boundary -- needs a real code block, not prose | 2506 | 1219 (-51%) | 3012 (+20%) | 3819 (+52%) |
| Average, all 20 prompts | 1268 | 946 (-25%) | 1081 (-15%) | 823 (-35%) |
Caveman's average looks better on paper than Atelier's full runtime (42% vs. 29% fewer output tokens) -- the honest number is the spread, not the average. Across 20 prompts, caveman's stdev is 30pp against Atelier's 21pp, and on error-boundary -- a prompt that actually needs a real code block, not prose -- the terse instruction backfires and produces 52% more output than baseline. A wording instruction can't tell the difference between a prompt that needs compression and one that needs a complete answer; it compresses everything, including the times it shouldn't. Atelier's number is smaller on average and far more consistent because it changes what gets retrieved and reused, not just how the reply is phrased -- it never regresses on 19 of these 20 prompts. Neither arm here touches turns, tool-call precision, or cost via caching; that's the full 5-rep, 200-run comparison in the flagship page.
The true story
Nobody publishes a benchmark for "just ask it to be terse" because it isn't a product -- it's the thing everyone already tries first. We ran it anyway, on the same prompts, same model, same reps as everything else on this site. Raw data, all 20 prompts, all 4 arms →