Claude-kit catalog-formatting

Overview

This skill covers the format and quality standards for the two catalog output files:

catalog.md

(human-readable) and

catalog.json

(machine-readable). The central quality concern is description writing — what a 2–3 sentence description must contain and what makes one good vs. bad.

Does not cover how to query the knowledge graph to retrieve papers (see

knowledge-graph-patterns

) or how to run catalog generation (invoke

catalog-generator

).

---

Description Writing

Every paper entry must have a 2–3 sentence description. This is the most important element of the catalog — it is what makes the catalog useful rather than just a list of links.

Required content per sentence

• Sentence 1 — the problem: what challenge or gap does this paper address? Why does it matter?
• Sentence 2 — the approach: what method, architecture, or technique does the paper propose?
• Sentence 3 — the result: what was achieved? Include numbers where the paper reports them.

Three sentences is the target. Two is acceptable when the paper is narrow. More than three is too long — the description becomes a summary, not a catalog entry.

Quality test

Read the description and ask: could someone decide whether to read the full paper based on this alone? If yes, it passes. If it is too vague to make that judgment, rewrite it.

Good description

Addresses the challenge of maintaining coherent long-horizon plans in LLM-based agents, where standard context windows fail beyond ~8K tokens. Proposes a hierarchical compression scheme that distills older context into semantic summaries while preserving recent tokens verbatim, implemented as a plug-in memory module. Achieves a 34% improvement on LongBench over standard RAG baselines while reducing KV cache size by 60%.

Why it works: specific problem, specific mechanism, specific numbers.

Bad description (too vague)

This paper presents a new approach to memory in language model agents. The proposed method shows strong performance on standard benchmarks.

Why it fails: does not say what approach, what benchmarks, what performance. Reader cannot decide whether to read the paper.

Common mistakes

• Opening with "This paper presents..." — start with the problem instead
• Citing relative improvements without the absolute baseline ("improves by 5%" — compared to what?)
• Using words like "novel", "state-of-the-art", "significant" without evidence
• Describing what the paper does instead of what it found

---

catalog.md Format

File structure

# Research Catalog: [Topic]

Generated: YYYY-MM-DD | Papers: N | Themes: M

## Table of Contents

- [Theme 1](#theme-1) (N papers)
- [Theme 2](#theme-2) (N papers)
- [Other](#other) (N papers)

---

<details>
<summary><strong>Theme 1</strong> (N papers)</summary>

[1-2 sentence synthesis: what this theme covers, what the papers in it collectively show]

| Paper | Description | Source |
|-------|-------------|--------|
| [**Title**](source_url) | [2-3 sentence description] | [![arXiv](https://img.shields.io/badge/arXiv-XXXX.XXXXX-b31b1b.svg)](https://arxiv.org/abs/XXXX.XXXXX) |
| [**Title**](source_url) | [2-3 sentence description] | [Semantic Scholar] |

</details>

---

<details>
<summary><strong>Theme 2</strong> (N papers)</summary>
...
</details>

---

<details>
<summary><strong>Other</strong> (N papers)</summary>

Papers that do not fit cleanly into any of the main themes.

| Paper | Description | Source |
|-------|-------------|--------|
...

</details>

Section synthesis

The 1–2 sentence synthesis at the top of each themed section should:

• Describe what the cluster of papers collectively addresses
• Note any shared methodology or point of convergence across the papers
• Not list individual papers — that is what the table is for

Theme naming

• Use noun phrases that describe the intellectual focus: "Agent Memory Mechanisms", "Multi-Agent Coordination Protocols", "Retrieval-Augmented Generation"
• Avoid generic labels: "Related Work", "Background", "Other Approaches"
• Maximum 4 words in the theme name

Paper ordering within sections

Order by citation count descending. For papers with zero citations (new preprints), order by date descending within that group.

---

arXiv Badge

For papers with an arXiv ID, use the standard badge format:

[![arXiv](https://img.shields.io/badge/arXiv-XXXX.XXXXX-b31b1b.svg)](https://arxiv.org/abs/XXXX.XXXXX)

Replace

XXXX.XXXXX

with the arXiv ID (e.g.

2401.12345

).

For papers without arXiv IDs, use the source name as plain text in the Source column:

• [Semantic Scholar]

  — link to Semantic Scholar paper page

• [PDF]

  — link to the fetched PDF

• [GitHub]

  — for repos and code papers

---

catalog.json Schema

{
  "generated_at": "2026-03-26T14:00:00Z",
  "topic": "Research topic string",
  "total_papers": 47,
  "themes": [
    {
      "name": "Agent Memory Mechanisms",
      "paper_count": 12,
      "synthesis": "1-2 sentence synthesis text",
      "papers": [
        {
          "id": "2401.12345",
          "title": "Exact paper title",
          "source": "arxiv",
          "source_url": "https://arxiv.org/abs/2401.12345",
          "arxiv_id": "2401.12345",
          "doi": null,
          "authors": ["Author One", "Author Two"],
          "date": "2024-01",
          "venue": "arXiv preprint",
          "citation_count": 342,
          "methodology_type": "empirical",
          "overall_assessment": "Strong",
          "code_available": true,
          "description": "2-3 sentence catalog description",
          "themes": ["Agent Memory Mechanisms", "Retrieval-Augmented Generation"],
          "arxiv_categories": ["cs.AI", "cs.LG"]
        }
      ]
    }
  ]
}

All fields are required. Use

null

for optional fields that are not available (doi, arxiv_id). Never omit a field.

---

When NOT to apply this skill

If you are querying the knowledge graph to retrieve papers for the catalog, use

knowledge-graph-patterns

. If you are running the full catalog generation pipeline, invoke the

catalog-generator

agent.