AI NewsDev toolsAnnouncement

Intent-Router writes a typed spec before a coding agent starts work

Intent-Router turns a short request to a coding agent into a typed IntentSpec file, asks one question only when the repo cannot answer it, and later checks the delivered code against the spec it wrote.

AI News

Editorial2 min read

LinkedInX
GitHub social preview card for the angel291592/Intent-Router repository

Image: GitHub

Why it mattersA team that keeps rewriting the same forty-question prompt before letting an agent touch the code can save the agreed version to disk and have the next session read it before probing anything.

A developer who writes 'add caching to the user API' to a coding agent usually gets a reply with another forty questions behind it. Intent-Router, an open-source skill by GitHub user angel291592, reads the repository first, finds what it can answer on its own, and only asks the one question the code and the docs do not already settle. The repository has 718 stars twenty-nine days after it was created and received a commit the day it reached the AI News candidate list.

What the file on disk looks like

The project README states that every run ends in a machine-readable file at .intent/<intent>.intent.yaml, written before the reply that references it. The file lists each contract clause, which fields were probed, where the evidence came from, and whether the intent is still waiting on an answer, settled, or already delivered. A later session reads that file before probing anything new. An answered question is not asked again. A settled contract is implemented as the file states.

One short measurement, from the project's own report

The project README links a comparison run from 24 September 2026 against the OpenCode harness. In that run, each of three deliveries written with Intent-Router stated the cache's failure policy. None of five bare deliveries did. The sample is three against five on one task, so the number is small and belongs to the project author rather than an independent test. The project README cites the report at evals/reports/2026-09-24-delivery-opencode.md for a reader who wants the raw run.

What it flags before the agent writes code

Intent-Router refuses to emit a spec while the intent is still underspecified. The project README lists four failure modes it names out loud: a request that contradicts itself, a request that reads two ways, a premise the repository already disproves, and a missing input the agent would otherwise invent. The agent asks for human judgment on the last of these and nothing else. The project's framing is 'prompt equity': the gap between the top one percent of prompt writers and everyone else stops deciding what comes back.

The limits of a 718-star project

Intent-Router is MIT-licensed and lists CI badges on its README. The 718-star number tells you other engineers found the idea interesting, not that a team of ten has run it in production for a quarter. The spec file format is the project's own, so a team that adopts the skill is also adopting that file as the contract its sessions read. The one documented delivery comparison is small and from the author.

Source

angel291592/Intent-Router on GitHub.

SourceGitHub

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project