What you get
The author wrote a whole page telling you when to switch it off.
Output Compression: 65% Is Prose, Not the Whole Story
Added 2026-08-06Six intensity levels, three of them classical-Chinese compression. A dedicated document says it plainly: 65% is the average over ten chat-style prompts, while a third-party run of 86 coding tasks measured 8.5%.
What deserves attention is when it loses money: the skill adds roughly 1,000–1,500 input tokens per turn, and if it saves less output than that you are paying to use it. The author names the losing cases: short coding exchanges, per-request billing, and one A/B where tokens multiplied instead — unreproduced but recorded anyway, concluding that wanting the rock to work is not the rock working. One rule runs against instinct: inventing abbreviations is banned, because tokenisers split them like full words.