grad-paragraph_skill

Overview

This skill produces a single, survey-quality paragraph that turns an evidence pack into a compact argument move following the tension → contrast → evaluation anchor → limitation micro-structure. It is designed to supply subsection-specific prose suitable for H3 body text without inventing facts or repeating template language. The output is conservative, evidence-bounded, and citation-embedded.

How this skill works

The skill reads the provided evidence snippets and planned argument lines, then emits one 4–6 sentence paragraph that opens with a focused tension, draws an explicit contrast between two clusters of approaches, names the evaluation anchor (benchmark/dataset/metric/protocol), and closes with a concrete limitation that affects transferability. It requires that citations are placed inside the sentences they support and that claims remain proportional to the granularity of available evidence (e.g., abstract-level vs experiment-level).

When to use it

• Drafting H3 subsection body text that must be evidence-grounded
• You have a complete evidence pack and a 4-line argument plan
• You want one compact paragraph rather than a multi-point list or template prose
• Preparing text for review where each paragraph must show an argument move

Best practices

• Create the 4-line Argument Planner first (tension / contrast / evaluation / limitation) and keep it out of the final paragraph
• Embed citations in the exact sentence they support rather than as a trailing list
• Use explicit contrast markers such as “whereas” or “in contrast” and avoid list-style comparisons
• Name the evaluation anchor (benchmark/dataset/metric/protocol) even at an abstract level to keep comparisons meaningful
• End with a specific, verifiable limitation tied to missing protocol details or incomparable measurements

Example use cases

• Summarize differences between retrieval-style and episodic memory designs with an evaluation anchor and limitation
• Compare tool-integrated agent pipelines versus simulator-only studies on a given benchmark metric
• Explain why two planning heuristics trade off success-rate versus budgeted API calls on a named dataset
• Turn scattered evidence notes into one concise paragraph for a literature-review subsection