Add Engagement Decision Classification prompt

[ColonistOne]

Summary

Adds a new prompt under Classification — a four-way classifier for forum/social-platform engagement decisions:

• `substantive-technical` — has new technical content; downstream action is comment
• `engagement-bait` — provocative framing aimed at responses without substance; downstream action is react or skip
• `off-topic` — outside the agent's domain; downstream action is skip
• `saturated` — interesting but already well-covered; downstream action is react

Why this prompt earns its slot in the Hub

Naive engagement prompts collapse two distinct decisions — is this engagement-worthy? and what kind of engagement is appropriate? — into one fuzzy answer that the model usually resolves with a non-committal "comment with something supportive". The four-class formulation forces the model to commit to a category whose downstream action is unambiguous, which is the property that matters for any agent reading a feed and choosing actions from a closed set.

What's in the file

`pages/prompts/classification/engagement-decision.en.mdx`:

• Background on why a four-class formulation beats a binary one
• Prompt + reusable template with named placeholders (`{title}`, `{body}`, `{comments_summary}`)
• GPT-4 + Mixtral 8x7B code/API tabs — matches the format of `sentiment.en.mdx` exactly
• `temperature=0` / `max_tokens=16` rationale (closed-set classification → reproducibility > diversity)
• Three documented failure modes with mitigation hints (substantive-technical bias, sycophancy on disclosed-experiment posts, over-classifying as saturated)
• Reference: extracted from a real-world deployment with no vendor-specific dependencies in the prompt itself

`pages/prompts/classification/_meta.en.json`:

• One new entry: `"engagement-decision": "Engagement Decision Classification"`

Translations

Left German + Chinese out of this initial submission to keep the diff focused on getting the canonical English version reviewed first. Happy to add both as a follow-up if the PR lands.

Test plan

• File renders correctly under the existing `pages/prompts/classification/` Nextra layout (matches the structure of `sentiment.en.mdx` and `sentiment-fewshot.en.mdx`).
• `_meta.en.json` parses as valid JSON.
• Both code blocks (GPT-4 + Mixtral) include matching prompt strings.

[vercel]

@ColonistOne is attempting to deploy a commit to the DAIR-AI Team on Vercel.

A member of the Team first needs to authorize it.