X Post Quality Extension Backend

X Post Quality Extension Backend

A browser extension can send the text of an X post to this workflow and receive a compact quality label. The extension decides how to display the result, for example by dimming low-value posts or highlighting useful ones. This template supplies the workflow backend, not the extension package.

Three labels

"AI slop" is a content-quality label here. It does not establish whether an AI wrote the post. A human can write filler and an AI-assisted post can be useful. The workflow judges only supplied text; it does not open links, verify claims, inspect images, or infer an author's intent.

What this workflow does

1. PostInput accepts a text field and optional post_id from the extension.
2. PostDecision asks a Choice question for the label and a Score question for information value, from 0 to 3.
3. ExtensionResponse returns a JSON object with label, raw_label, confidence, probabilities, information_value, review_recommended, and post_id.
4. If category confidence is below 0.65, the displayed label becomes neutral while the raw model choice remains available for inspection.

Setup

Select a Decision Model credential on PostDecision. Keep ExtensionResponse as the only terminal node so the execute endpoint can return its mapped JSON directly. The example model is jev-latest; choose a supported model for your endpoint.

Use the editor's Run with cURL dialog to get your workflow's execute URL and configure authentication. The endpoint is POST /api/workflows/{workflow_id}/execute. Send JSON such as:

{
  "text": "We reduced p95 query latency from 840 ms to 120 ms by replacing N+1 lookups with one indexed join. Here is the query plan and the benchmark method.",
  "post_id": "demo-post-001"
}

The example request is a post to classify, not a verified performance claim. A strong label still depends on what evidence the supplied text actually contains.

Connect the extension

Send the captured post text from the extension's background worker to your authenticated backend endpoint. Configure host permissions for that endpoint. Keep shared backend credentials on a server-side proxy instead of embedding them in a distributed extension; the extension should authenticate its user to that proxy. Provider credentials stay inside Heym.

The execute response uses the simple-response mode by default. Keep that mode enabled when consuming the mapper object directly; X-Simple-Response: false returns the execution envelope for debugging. Treat HTTP failures as an unavailable classification, not as an AI-slop label.

Try these posts

• AI slop: "This changes EVERYTHING. Ten secrets nobody tells you. Comment YES and follow for the ultimate productivity revolution."
• Golden post candidate: "To reproduce our bug: create an empty list, call pop(), and observe IndexError. Check the list before removing an item; here is the two-line fix."
• Neutral: "Enjoyed a quiet walk this morning. Hope everyone has a good weekend."

Tune the criteria and confidence threshold against posts your users have labeled. Use review_recommended to leave uncertain posts unchanged and the full probability distribution to explain the result in a tooltip.