Skip to content
BUILTLive

ROUTR

Small router skills that send an AI coding agent to the right workflow before it reads anything.

Snapshot

My role
Author and maintainer
Timeframe
July 2026 – present
Status
Live
Team size
1

The problem

An agent with dozens of installed skills tends to pick the wrong one and read whole files to find out, burning its context window. ROUTR puts a thin layer in front: 28 routr-* router skills with explicit “use when / not for” descriptions, so routing is decided from the slash-menu text and only the one full playbook needed gets loaded. Requests that span several workflows get a route plan of at most three steps, with one router active at a time.

How the project ran

I owned

  • Designed the router format: a tight “Use when / Not for” description, an exit condition and a handoff
  • Wrote the 28 router skills and the routes.json file that defines handoffs, companions and chains
  • Built the skill validator that runs in CI on every change
  • Wrote the 153-prompt single-router eval set plus multi-route and held-out chain prompts, and ran the blind evaluations

Constraints

  • Routing has to work from short descriptions alone — the agent should not load a playbook to decide whether it needs it
  • Real requests span several workflows, but two routers active at once would contradict each other
  • Child skills may not be installed, so every router needs a fallback

Outcome

  • MIT-licensed and installable with one command
  • Blind runs with a model that saw only the prompt file and menu reached 99.3% and 97.4% router accuracy on the 153-prompt set, as reported in the README; the method is published so the numbers can be reproduced
  • Route plans capped at three steps, with one active router and at most one companion checklist per step
  • Missing child skills fall back to built-in routr-depth-* skills instead of failing

Learned

  • Description text is the routing interface — the “not for” boundaries matter as much as the “use when” ones
  • Chain prompts written before the composition rules, and never used for tuning, give a more honest measure than a tuned set
What I owned, what constrained it, and where it got to.

Built with

  • Python
  • PowerShell
  • Shell
  • GitHub Actions

Evidence

StartupLink interface showing the founder and student onboarding flow
BUILTIn development

StartupLink

A closed network where students and verified student founders can find each other.

Role
Founder and full-stack developer — sole developer
Timeframe
April 2025 – present
BUILTIn development

Nudgio

An offline-first Android alarm that plays the user’s own media, with no backend and no Internet permission.

Role
Sole developer
Timeframe
August 2026 – present