01 · Public source

PromptCore

A local router for choosing a Prompt template before the model starts answering.

The routing decision is calculated locally. The downstream model receives the resulting Meta-Prompt together with the original task.

View source

Problem

Code, planning and research are different kinds of work. Using the same prompt logic for all three did not make sense to me.

Mechanism

Read the task and estimate its category, complexity and intent.
Rank 40 registered reasoning templates and select one of them.
Build a Meta-Prompt without calling a model; leave the downstream answer to the caller.

Evidence

Bundled tests: 22 / 22 passed
Local pipeline: 0.0430 ms p50 · 2,000 runs
Exact framework match: 19%
Synthetic coverage: 20 / 40 templates

Boundary

The benchmark uses author-defined expected templates, so it is an internal calibration tool rather than independent validation. The results show that routing is quick and reproducible; they do not show that the selected template produces a better downstream answer.

Notes

2026-08-26 Why prompt routing belongs outside the model