Honey aims to make AI coding answers much shorter
Honey is an prompt style for making give shorter answers. It combines Ponytail’s focus on writing only the needed code with Caveman’s very brief writing style, then adds ESO for passing information between AI agents. The benchmark used 23 tasks on , with 3 runs per task and a 4-model judge panel using a neutral scoring rule.
On 14 coding tasks, Honey kept quality at 98% while cutting tokens by 49%. On 7 user-facing writing tasks, it reached 101% quality with 6% fewer tokens. On 2 agent-to-agent handoff tasks, it kept 100% quality while cutting tokens by 51%.
Ponytail can become longer than expected because it always adds a self-check, even for simple coding tasks. Caveman can compress handoffs so much that important meaning becomes hard to recover, while Honey uses ESO instead of pretty JSON to keep the same structure in fewer words.
Key points
- Honey is a prompt style for , built from ideas in Ponytail and Caveman.
- In 14 coding tasks, it reported 98% quality with 49% fewer tokens.
- In user-facing writing tasks, it reported 101% quality with 6% fewer tokens.
- For agent-to-agent handoffs, ESO helped keep 100% quality while using 51% fewer tokens.
- It is and .