| name | exhaustive long-context coverage |
| description | Use this skill when the user wants questions that 'need reading everything', 'leave no file behind', 'spread the clues across many documents', 'list all entities associated with a topic', or 'test whether the model skips scattered fragments'. Trigger it for large multi-document bundles where the answer depends on broad coverage, assembling high-cardinality lists, or piecing together small discontinuous facts spread across massive reports. |
Skill: exhaustive long-context coverage
1. Capability Definition & Real Case
- Professional Definition: The ability to preserve document-level coverage over extremely long multi-document contexts or dense structured arrays, ensuring no required document or entity is silently ignored. The core challenge involves broad coverage discipline, high-cardinality entity tracking, and bridging physically disconnected but semantically related token fragments.
- Dimension Hierarchy: Multi-Source Evidence Composition->Long-Context Coverage->exhaustive long-context coverage
Real Case
[Case 1]
- Initial Environment: A long bounded bundle of company reports from the same domain. Each company’s non-current-assets figure is stored in its own document, and the answer requires scanning all candidate companies before comparing them.
- Real Question: Which company has the highest non-current assets?
- Real Trajectory: 1. Map all documents in the bundle to their respective companies. 2. Sequentially extract the non-current-assets figure from each document. 3. Compare the complete extracted array to find the maximum value. 4. Output the company matching the maximum value.
- Real Answer: HARTE HANKS
- Why this demonstrates the capability: This case punishes the agent if it reads only a convenient subset of documents. The answer is only trustworthy after coverage across all comparison candidates. The capability is therefore exhaustive long-context accounting, not simple local retrieval.
[Case 2]
- Initial Environment: A large collection of short scholarly biography documents, each describing the career of a single award winner, including their primary research area and nationality.
- Real Question: What is the distribution of all Turing Award winners by their fields of study by 2023?
- Real Trajectory: 1. Scan the entire document set to build a master list of all identified Turing Award winners. 2. For each identified person, locate and extract the specific field of study mentioned. 3. Aggregate the individual field assignments. 4. Synthesize the final comprehensive distribution logic.
- Real Answer: As of 2023, the Turing Award winners are distributed across various fields: Artificial Intelligence (X winners), Cryptography (Y winners), Database Systems (Z winners)... [Full exhaustive list provided].
- Why this demonstrates the capability: The answer requires exhaustive entity recall across dozens of documents. Failure to read even one document or misattributing a field to the wrong winner degrades the accuracy of the statistical distribution.
[Case 3]
- Initial Environment: A RAG context containing a massive 75-page Advanced Persistent Threat (APT) report detailing an operating sequence. The tool used for execution and the resulting file suffix are separated by thirty pages of irrelevant log data.
- Real Question: What is the specific command-line argument used for data compression in the observed exfiltration sequence, and what file extension does the resulting archive use?
- Real Trajectory: 1. Scan the 75-page report for data collection parameters. 2. Extract the '-mx9' command argument on page 14. 3. Skip extensive background narratives. 4. Locate the discontinuous token span on page 19 describing the '.tmp.enc' suffix. 5. Merge the fragmented pieces into a localized whole.
- Real Answer: The command-line argument used for compression is '-mx9' and the resulting exfiltration files use the '.tmp.enc' extension.
- Why this demonstrates the capability: The relevant evidence is completely discontinuous. The specific facts are separated by dozens of irrelevant paragraphs, requiring sub-document tracking and multi-span bridging without summarizing the useless middle noise.
Pipeline Execution Instructions
To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:
-
Phase 1: Environment Exploration
Read the exploration guidelines to discover raw knowledge seeds:
references/EXPLORATION.md
-
Phase 2: Trajectory Selection
Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
references/SELECTION.md
-
Phase 3: Data Synthesis
Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
references/SYNTHESIS.md
