agents2agentsai/cross-paper-report

Cross-Paper Report

For conversation-derived reports, use $conversation-report. For LaTeX PDF output, use $latex-report after generating the narrative.

You MUST produce a deep narrative explanation — no exceptions, no shortcuts. A short summary is NEVER acceptable.

Word-Count Floors

Hard minimums. If any section falls below, STOP and expand before proceeding.

| Tier | Word floor | When to use | |------|-----------|-------------| | Focal | 800–2500 words | 1–4 cards (all focal), or the 2–4 most novel/complex/central in a 5–12 set | | Supporting | 400–600 words | Remaining papers in a 5–12 set |

Core Principles

Three principles govern all output from this skill:

1. TTS-readable — Output is read aloud. Keep paragraphs short, pacing natural. No walls of text.
2. Teach, don't summarize — The reader has not read the paper. They should understand the method after reading your explanation. Ground everything in concrete numbers from the paper. Explain every equation intuitively. Establish why before how.
3. Self-contained sections — Each section in the reading view loads independently. Never forward-reference ("as we'll see in Section 3"). Explain concepts completely where they appear. Define every term on first use.

Hard Rules

• No academic citations. Never "Name et al. (YYYY)". Use paper titles or natural phrasing ("the LAPA paper showed…"). TTS reads these aloud.
• Never reference the KB. Present content as if you understand the papers directly.
• Details block at the end of each subsection: 1–3 sentences collecting model names, dimensions, hyperparameters. The narrative must be understandable without it.
• Voice: Second-person for procedural walkthroughs ("You take two frames…"). Neutral third-person for framing and results. Do NOT open problem statements with "You want to…".

Prerequisites

Determine <kb_path> per $kb (default ${CODEX_KB_PATH}). If KB is disabled, work from card content in the conversation. Skip all KB reads/writes.

Scale Detection

After reading all requested cards:

• ≤ 12 cards: Follow Phases 0–2 below.
• > 12 cards: Jump to Large-Set Synthesis at the end.

Phase 0: Related Card Discovery

Skip if KB is disabled or running as a cluster subagent.

1. List all cards per $kb. Scan tags/connections for related unrequested cards.
2. If found, present them: "Your KB also has cards for [list] — want to include any?"
3. Auto-include without asking on "full survey" / "explain them all", $paper-discovery pipeline, or auto-continue flow. Skip entirely if user says "only these cards."

Phase 1: Deep Technical Explanation

Read all cards per $kb. If KB is disabled, use conversation context.

Coverage-First

Never drop cards. If 10 are relevant, all 10 appear. Use tiering or Large-Set Synthesis — never cut.

Tiered Depth

• 1–4 cards: All focal.
• 5–12 cards: 2–4 focal (most novel/complex/central), rest supporting.
• 13+: Large-Set Synthesis.

Cluster subagents: all cards focal, skip user confirmation. When user confirmed scope or triggered from $paper-discovery, proceed without asking.

Focal Card Elements

Organize in whatever sequence tells the clearest story. Integrate equations, analogies, and Details blocks where they naturally belong:

• Problem (plain-language gap), architecture, training pipeline, equations with intuition, results with specific numbers and baselines, limitations

Supporting Card Structure

5 substantive paragraphs: problem & core idea, method, key technical detail, results with numbers, connection to focal cards. No stage walkthroughs or Details blocks needed.

Phase 2: Cross-Card Comparative Synthesis

Produce when 2+ cards. Skip for single-card requests. Read research-context.md (if KB enabled) to frame around user priorities.

Compare along concrete technical dimensions with specific numbers and implementation details. Close with a How the Papers Relate prose section.

Before completing: verify word-count floors are met. If any section is below, expand it.

Presentation (Main Agent Only)

Cluster subagents never call present_reading_view — return content as text.

Section length: Each section ≤ 40 lines. Focal papers split into 2–3 ## sub-sections (Problem & Core Idea, Method, Results & Analysis). Supporting papers: single ## section each. Comparison: ## How the Papers Relate, split if longer.

Phase 1 (Outline): IMMEDIATELY call present_reading_view with document_id, title, and content containing ONLY ## headings with empty bodies.

Phase 2 (Fill): The tool result tells you the next section. Call update_document_section for each sequentially.

Markdown: Always put a blank line before list items — without it, the parser treats them as plain text.

Follow-ups: append_to_section with foldable=true for expansions, patch_document_section for corrections, update_document_section only for full rewrites. Keep sections ≤40 visible lines. No editorial labels in headings.

Post-Report Housekeeping

Skip if KB is disabled.

1. Journal — Prepend to <kb_path>/research-journal.md: date, title, papers compared, key findings (1-2 bullets), card IDs touched.

2. Research context — If user shows preference signals during Q&A, briefly offer to note in research-context.md. Never block the interaction.

3. Follow-up persistence — When Q&A produces new insights, spawn fire-and-forget $kb subagent to persist them (card IDs + new insights, 2-4 sentences each).

---

Large-Set Synthesis (> 12 cards)

Step 1: Cluster Assignment

Group into 3–6 thematic clusters (4–8 cards each). Present to user: cluster names with rationale, card IDs, bridge cards, and a shared terminology glossary (5–10 cross-cluster terms).

Step 2: Per-Cluster Sub-Reports

Launch one subagent per cluster in parallel. Overrides: skip Phase 0, skip Presentation (return text), all cards focal, no user confirmation, include shared glossary.

Each returns: per-card walkthroughs + within-cluster synthesis, cluster summary, cross-cluster interface (key terms, equations, specific numbers, bridge points).

If a subagent fails, proceed with remaining. Note the gap, offer retry.

Step 3: Meta-Synthesis

New analytical layer connecting clusters (not a summary). Required sections: Scope & Framing, Cross-Cluster Evolution Tracing, Design Space Mapping, Failure Mode Lineage, Key Equations Across the Stack, How the Papers Relate, Practical Takeaways. Use cross-cluster interfaces as primary data source. Ground in specific numbers.