picooWhat smaller tool prompts change
Picoo searches for one operation before asking a model to call it. We tested whether that smaller prompt changes token use, success, and speed.
Main result
Less context. More successful calls.
Token use
The prompt made the difference
Output token use stayed close. Input token use fell by 86.0% because Picoo showed the model one selected contract instead of all 13 tool schemas.
Method
What we tested
We used deterministic tasks so every condition had the same expected operation and arguments.
- 01Matched tasks
Each task had one correct operation and one correct set of arguments.
- 02Two prompt shapes
Picoo received one selected contract. The comparison received all 13 tool schemas.
- 03Strict success
A call passed only when both the operation and every argument were correct.
Findings
What we learned
Most savings came from input. Picoo used 258,378 input tokens. The full registry prompt used 1,844,989.
Tool selection improved. Picoo selected the correct operation on 89.8% of tasks. The full registry prompt reached 35.8%.
The token result repeated. Picoo used fewer tokens for every model in the eight model comparison. Reductions ranged from 70.8% to 83.7%.
Speed depended on the model. Picoo was faster on six of eight models and slower on two. A smaller prompt does not guarantee a faster final response.
Limits
What this does not prove
- The large task set was generated. It was not private production traffic.
- The full registry condition used MCP shaped schemas, but it did not call a live MCP server.
- The 1,000 task run used one model on one local Ollama host.
- We have not yet compared Picoo with a modern MCP client that loads tools only when needed.
The evidence supports one clear claim: when the alternative puts a full tool registry in the prompt, selecting one small contract first uses far fewer tokens.