decompose-linear-multi-hop-dependency
$
npx mdskill add Memento-Teams/Memento-Widesearch/decompose-linear-multi-hop-dependencyDecomposes sequential queries where each answer is required as the search key for the next step in a chain.
- Handles queries with nested relationships or possessive chains requiring multiple linked searches.
- No specific tools or APIs mentioned; likely relies on general search or reasoning capabilities.
- Uses a four-step template to identify entities, extract attributes, apply constraints, and verify results.
- Presents results through a sequential chain-of-thought process without parallelization.
SKILL.md
.github/skills/decompose-linear-multi-hop-dependencyView on GitHub ↗
--- name: decompose-linear-multi-hop-dependency description: Decomposes queries where each answer serves as the mandatory search key for the subsequent step in a sequential chain. --- ## When to Use Use this skill when the query contains a "possessive chain" or a series of nested relationships. This is indicated by phrases like "the [Attribute] of the [Entity] who [Action]," or when the final answer depends on a specific property of an intermediate entity that is not mentioned in the original prompt. If the query requires identifying Entity A to find Entity B, and Entity B to find Entity C, it is a linear multi-hop dependency. ## Decomposition Template 1. **Step 1: Identify Anchor Entity.** Extract the primary known entity and the specific constraint (date, location, or title) to find the first hidden link. 2. **Step 2: Extract Bridge Attribute.** Using the result from Step 1, search for the specific property or relative required to move to the next hop. 3. **Step 3: Final Constraint Application.** Using the result from Step 2, apply the final filter or count requested in the original prompt. 4. **Step 4: Verification.** Re-trace the chain from Step 3 back to Step 1 to ensure no "entity drift" occurred (e.g., ensuring a date constraint from Step 1 still applies to the final result). ## Worker Assignment Rules - **Single Worker Execution:** Assign to a single "Chain-of-Thought" worker. Do not parallelize, as Step N cannot begin without the output of Step N-1. - **Sequential Verification:** If the chain exceeds three hops, insert a verification step after the second hop to confirm the "Verified Entity" before proceeding. ## Answer Format The output must show the reasoning chain for each hop. End with: `最终答案:你的答案` ## Anti-Patterns - **Premature Parallelization:** Attempting to search for the final entity and the bridge entity simultaneously, leading to "hallucinated" connections. - **Constraint Leaking:** Forgetting a specific year or numerical constraint mentioned in the first hop by the time the agent reaches the final hop. - **Entity Ambiguity:** Failing to resolve a common name in the middle of the chain, causing the search to pivot to a different person/object with the same name. - **Counting Logic Errors:** Confusing "number of occurrences" or "number of terms" with "number of unique individuals" when the final hop involves a list.