Run3 product grounded artifact retrieval
Load enterprise data from /root/DATA, identify the correct artifact version for each product mentioned in questions, and reject cross-product distractors. Apply strict 2-signal product grounding (artifact metadata + question context).From its SKILL.md
npx -y skills add cxcscmu/SkillLearnBench --skill run3_product-grounded-artifact-retrievalAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
SKILL.md
1.5 KB, 287 tokens by cl100k_base, as published. Nobody here has run it
Product-Grounded Artifact Retrieval Skill
Step 1: Load Enterprise Data
- Read all files from
/root/DATAdirectory - Index artifacts by product name and version (draft vs. final)
- Track artifact metadata: creation date, status, product association
- Build a product-to-artifacts mapping
Step 2: Two-Signal Product Grounding
For each question's artifact context:
- Signal 1: Match product name from question to artifact metadata
- Signal 2: Verify artifact status/version (prefer final versions unless explicitly asking for draft)
- Reject any artifacts from different products, even if semantically similar
- Reject draft versions unless the question specifically asks for drafts
- Document which artifacts were rejected and why
Step 3: Validate Product Isolation
- Ensure no cross-product confusion (e.g., Product A reviewers mixed with Product B)
- If multiple artifact versions exist for a product, select the correct one based on question intent
- Fail explicitly if product grounding is ambiguous
Step 4: Return Grounded Artifacts
Return a mapping of:
{
"question_id": "q1",
"artifact_id": "artifact_name",
"product": "product_name",
"version": "final|draft",
"valid": true/false,
"rejection_reason": "if not valid"
}
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.