ModerationPrompts

System and user prompts used for LLM-based content moderation. Each prompt instructs a vision-language model to classify a social-media post (text + image(s)) against a safety policy and return a structured JSON decision.

---

Repository Structure

prompts/
  bluesky_policy/           ← prompts grounded in Bluesky community guidelines
    community_guidelines.md     policy document (included verbatim in some prompts)
    prompt_labels.md            
    prompt_labels_details.md    
    prompt_labels_rationale.md  
    prompt_labels_rationale_details.md  
    prompt_labels_fewshot.md           few-shot variant — labelled examples before target post
  meta_policy/              ← prompts grounded in Meta content policy
    all_category_w_explanation.md  all harm categories with detailed explanations

---

Bluesky-Policy Prompts

File	Includes community guidelines	Label detail level
community_guidelines.md	— (guidelines doc only)	—
prompt_labels.md	No	Basic
prompt_labels_details.md	No	Detailed (with sub-rules)
prompt_labels_rationale.md	Yes	Basic
prompt_labels_rationale_details.md	Yes	Detailed (with sub-rules)
prompt_labels_fewshot.md	No	Basic + few-shot examples

---

Meta-Policy Prompts

File	Policy source	All harm categories	Includes explanations	Requests rationale
all_category_w_explanation.md	Meta community standards	Yes	Yes	Yes

all_category_w_explanation.md — full prompt grounded in Meta content standards. Covers all harm categories with per-category explanations.

---

Prompt Format

Each .md file contains clearly delimited sections:

SYSTEM PROMPT:
<system instruction — role, task, output format constraint>

------------

USER PROMPT:
# **COMMUNITY GUIDELINES**   ← present in _rationale and meta_policy variants
...

# **LABELING POLICY DETAILS**
## **S1**: porn
...

POST CONTENT:
<post text and image placeholder>

OUTPUT FORMAT:
<required JSON schema>

---

License

CC BY-NC 4.0 — non-commercial use only.