rsshub-route-authoring-skill

An agent skill for authoring PR-ready RSSHub routes — code that survives code review, not just code that runs.

English | 中文

This is an unofficial community project. It is not affiliated with, endorsed by, or connected to RSSHub or its maintainers. It is a summary of publicly visible review feedback, offered as advice only: whenever it disagrees with AGENTS.md, the official docs, or a maintainer, they are right and this is wrong.

Heads up: this skill has a feedback loop, and it never runs on its own

If an agent uses this skill and a reviewer raises something it did not cover, the skill may offer to report that back here. It will never do so unless you explicitly agree.

Concretely, the skill is instructed to:

• ask first, every single time — no exceptions;
• never open an issue, open a pull request, edit a file, or make any network call about this repository unless you asked it to;
• if you decline, drop the matter entirely.

If you do agree, the report is filed with your GitHub account via gh, which means it appears publicly under your name. Nothing leaves your machine otherwise.

This matters because the skill is built from real review feedback — but a feedback loop that runs silently would be a betrayal of the person who installed it. So: visible, opt-in, and always under your control. Details in CONTRIBUTING.md.

Provenance of the writing

The rule content here is derived from real review data and is reproducible with tools/. The explanatory prose was drafted with substantial AI assistance and reviewed by a human contributor. That combination has a particular failure mode worth stating plainly: AI-written guidance tends to state patterns more confidently than the evidence supports.

Two consequences for readers:

• Every count in this repository can be recomputed with npm run fetch && npm run analyze.
• Any claim that carries neither a count nor a rule ID should be treated as unverified.

Most RSSHub route guides explain how to fetch and parse a page. That part is easy. The hard part is the review. This skill is built from the review record itself, so that an agent can get it right the first time and not spend a maintainer's evening on avoidable round trips.

Thanks

RSSHub is maintained by volunteers who review an extraordinary volume of contributions. Every number in this repository is a small window onto work they do for free. Nothing here is intended as a way to game review — the goal is the opposite: send PRs that need less of their time, and learn from what they have already had to explain hundreds of times.

The evidence

Every claim traces back to real reviewer feedback, not to intuition:

Corpus	2,727 route PRs from DIYgod/RSSHub — 1,987 merged, 740 closed without merge
Comments	18,680 (5,037 human, 589 auto-review, 13,054 other bot)
Collection	GitHub GraphQL via gh; feat(route) and fix(route), each split into merged and closed-without-merge
Reproduce	npm i && npm run fetch && npm run analyze

The pipeline lives in tools/ and is re-runnable, so the counts can be refreshed as review practice changes. See the caveat at the bottom about what this data can and cannot tell you.

What reviewers most often have to ask for

Ranked by how ofte