agentsclimarketplace

Run3 product grounded artifact retrieval

Skill cxcscmu/SkillLearnBench/skills/b3-teacher-feedback-claude-haiku-4-5/enterprise-information-search/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

Install
npx -y skills add cxcscmu/SkillLearnBench --skill run3_product-grounded-artifact-retrieval

Assembled 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/DATA directory
  • 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.

Keep looking

Skills are one crate of 326,852. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.