agents2agentsai/research-briefing

Research Briefing

A 2-4 page orientation document that cuts through paper complexity to extract the one core idea per paper. Answers "what are my options and which should I look at first?" — not "how does every mechanism work in detail."

For full technical walkthroughs, use $cross-paper-report. For conversation-derived reports, use $conversation-report. This skill is for quick orientation before diving deep.

Core Principle: Cut Through the Noise

Most papers wrap a small, concrete idea in 15 pages of related work, ablations, and notation. The briefing extracts that one core idea and states it plainly — not in the paper's language, but in yours.

Rules:

• State the core contribution in 1-2 sentences using plain language.
• If the idea is small, say it's small: "This paper's contribution is a single architectural change: replacing the action head with a diffusion decoder."
• If it's incremental, say what it's incremental over: "This is RT-1 but with a VLM backbone instead of a ViT — same tokenized action approach, bigger pretrained model."
• Never mirror the paper's framing of its own importance — papers are designed to sound maximally novel.
• Use the pattern: "[Paper] does [one concrete thing]. The trick is [specific mechanism]. Tradeoff: [what you gain vs. what you lose]."
• Per-paper depth: 3-5 sentences max. Must include the one core idea (plainly stated), one specific number or result, how it relates to the other approaches, and what it trades off.
• Citation formatting: NEVER use academic citation style like "Smith et al. (2026)" or "Author (Year)" — reading view content is read aloud and academic citations sound unnatural when spoken. Instead, reference papers by title: "the LAPA paper showed..." or use natural descriptions: "researchers found...", "a recent study demonstrated...". Never put DOIs or arXiv IDs inline in paragraphs. Collect full references (with IDs) in a References section at the end of the briefing if the user needs them for follow-up. When the method name or paper title is short and recognizable, prefer it over a generic phrase — say "the GRPO method showed" or "the Attention paper demonstrated" rather than just "researchers showed". Use "researchers showed" only when neither the method nor the paper name is short or memorable.
• Never reference the KB in explanations. Do not say "as summarized in your KB" or "your KB card says." The KB is infrastructure — present content as if you understand the papers directly.

Phase 0: Source Material

This skill works best with KB cards, but it can also work without them.

Path A: KB Available (preferred)

1. Determine <kb_path> per the $kb skill and verify KB exists.
2. List all cards per $kb.
3. Filter cards whose tags, titles, or topics relate to the user's question or requested topic.
4. If related unrequested cards exist, present them: "I also found cards for X, Y — want me to include them in the briefing?"
5. If no cards match the topic but KB is enabled, tell the user: "No KB cards found for this topic. Run $paper-synthesis on relevant papers first, then re-run this briefing." The briefing synthesizes from existing KB content — it does not read papers itself.

Path B: No KB, but conversation context available

If KB is disabled but the conversation already contains synthesis output (from $paper-synthesis staging files, prior explanations, or user-provided paper details), use that content as source material. Extract the same fields as Phase 1 (core contribution, key result, approach family, tradeoffs, status) from whatever is available.

Path C: No KB, no prior context (cold start)

If KB is disabled and no prior synthesis exists in the conversation, run a lightweight discovery:

1. Use paper_search with 2-3 facets from the user's topic (same facet design as $paper-discovery Phase 1).
2. Fetch the top 8-12 results sorted by citation count.
3. Use the paper abstracts and metadata as source material for the briefing — this produces a shallower briefing than Path A, but still gives the user actionable orientation.
4. Note in the briefing: "This briefing is based on paper abstracts only. For deeper analysis, run $paper-synthesis on papers of interest."

Path C is a fallback. The briefing will lack the depth of a KB-sourced briefing — abstracts don't contain method details, specific numbers, or implementation insights. But it still answers "what are my options?" and gives the user a starting point.

Phase 1: Read and Analyze Source Material

For each relevant card (or paper abstract in Path C):

1. Read the card per $kb to get the full content. For Path B, read from conversation context. For Path C, use the paper_search results directly.
2. Extract:
   • The core contribution (from Summary or Deep Dive — the single most important idea)
   • The key result (one specific number with context)
   • The approach family (what paradigm does this belong to?)
   • Tradeoffs (what does this approach gain and lose?)
   • Status (current state of the art? Superseded? Foundational?)

Phase 2: Cluster into Approaches

Group the papers into 2-5 approach families based on their core paradigm. Each approach should have a clear name and a 1-sentence description of the shared strategy.

Phase 2.5: Read Research Context (Optional, skip if KB is disabled)

If <kb_path>/research-context.md exists, use it to tailor the briefing:

• Priorities: Weight the Recommendation section toward the user's documented priorities (e.g., if they care about inference latency, lead with approaches that address it).
• Not Interested In: De-emphasize or briefly note approaches the user has dismissed, rather than giving them equal airtime.
• Framings That Work: Use explanation styles that match the user's preferences (e.g., tradeoff framing, concrete numbers).

If the file doesn't exist, skip this step — the briefing still works without it.

Phase 3: Generate the Briefing

The briefing is delivered in the chat response only — presented directly in conversation. This is the sole deliverable. No markdown file, no PDF. A briefing is a quick orientation — persistence adds latency without value. If the user wants to re-read a previous briefing, they can re-run it (it reads from KB cards, which are stable). If the user wants a polished typeset document, they should use $cross-paper-report.

Research journal tip: If the user has a research-journal.md in their KB, previous journal entries may show related topics they've already explored — mention this connection if relevant (e.g., "You explored action tokenization on [date] — this briefing covers the broader landscape around that topic").

Briefing Structure

## Research Briefing: [Topic]
### [Date]

### The Landscape
[1 paragraph: what problem space this covers, what the main tension is,
where the field is heading. 4-6 sentences max.]

### Approaches at a Glance

#### [Approach 1 Name] (N papers)
**Core idea:** [1-2 sentences — the actual contribution, plainly stated]
**Key paper:** [Title] ([Year]) — [why this is the best representative]
**Tradeoff:** [what you gain vs. what you lose in 1 sentence]
**Best for:** [when to use this approach]

#### [Approach 2 Name] (N papers)
**Core idea:** [1-2 sentences]
**Key paper:** [Title] ([Year]) — [why this is the best representative]
**Tradeoff:** [what you gain vs. what you lose]
**Best for:** [when to use this approach]

...

### Per-Paper Quick Reference

| Paper | Year | Core Idea (1 sentence) | Key Number | Status |
|-------|------|----------------------|------------|--------|
| RT-2 | 2023 | VLM predicts tokenized actions directly | 2x generalization on novel objects | Superseded by OpenVLA |
| OpenVLA | 2024 | Open 7B VLA with LoRA adaptation | 16.5% avg improvement with fine-tuning | Current best bootstrap |
| ... | | | | |

### Recommendation
[1-2 paragraphs: Given what you're trying to do, here's what I'd suggest
looking at first and why. Reference specific papers. Be opinionated —
the whole point of a briefing is to help the user decide where to focus.]

### Dive Deeper
For full technical walkthroughs of any paper above:
- `$cross-paper-report [card-ids]` for deep multi-paper analysis
- `$paper-synthesis [arxiv-url]` for single-paper deep dive
- Chat about any paper and ask follow-up questions — insights will be saved to KB cards per `$kb`

Presentation (Main Agent Only)

This section applies to the main agent only. If this skill is loaded in a subagent, return results as text to the main agent — do NOT call present_reading_view.

Phase 1 (Outline): IMMEDIATELY call present_reading_view with document_id set to a unique slug, title to the briefing title, and content containing ONLY the ## section headings with empty bodies. Example content: "## The Landscape\n\n## Approaches at a Glance\n\n## Recommendation". This opens the reading view instantly with "Generating..." placeholders.

Phase 2 (Fill): The tool result will tell you to fill section 0. Immediately call update_document_section(document_id, section_index=0, content="...") with the FULL content for that section — do not output any text, just make the tool call. Each tool result tells you the next section to fill. Continue calling update_document_section for each subsequent section until all are filled.

Markdown formatting: Always put a blank line before numbered list items (1., 2., etc.) and before bullet list items (-, *). Without a blank line, the markdown parser treats 2., 3., etc. as plain text instead of list items, so they lose their formatting. This also applies to content after paragraphs, blockquotes, and code blocks.

When the user asks follow-up questions about a specific section, use the most efficient update tool:

• append_to_section with foldable=true — preferred for expansion requests (e.g., "explain more", "go deeper on X"). Adds a collapsible detail block below existing content, preserving scannability. Each foldable block: 3-5 sentences.
• append_to_section (without foldable) — to add genuinely new information at the end of a section.
• patch_document_section — to change specific text within a section (for corrections or targeted edits).
• update_document_section — to fully rewrite a section ONLY when the user asks for a different framing or the section is factually wrong. Never use this just to add more detail.

Critical constraint: After any follow-up update, the section must still be ≤30 lines of visible (non-folded) content. If a request would push past this limit, use foldable blocks.

Write follow-up answers as straight content — no editorial labels like "(clearer explanation)" or "(expanded)" in headings or topic lines.

Post-Briefing Housekeeping

Skip this entire section if KB is disabled. When KB is off, do not write journal entries, research-context updates, or spawn $kb subagents. The briefing in the reading view is the sole output.

After presenting the briefing, do these:

1. Journal entry — Append to <kb_path>/research-journal.md. Prepend (newest first):

## [Date] — Briefing: [Topic]

### Explored
- Briefed on [N] papers covering [topic]
- Approaches identified: [list approach families briefly]

### Recommendation
- [1-2 sentences: what was recommended and why]

### Cards Touched
- [card-ids] (read)

---

2. Research context detection — If the user's questions or reactions reveal priorities (e.g., "which of these is fastest at inference?", "I don't need the simulation-only ones"), offer to note it in research-context.md. This is especially valuable during briefings because the user is actively deciding what to focus on.

3. Follow-up Q&A persistence — When the user exits the reading view and Q&A produced new insights not already in the relevant KB cards, automatically spawn a fire-and-forget $kb subagent (do NOT call wait) to persist them. Do not ask the user. Include the card IDs and a summary of new insights in the subagent prompt:

$kb

Update KB cards with follow-up insights. Do NOT ask the user — this is automatic. Cards to update: [card-id-1], [card-id-2], ... New insights from Q&A: [For each card: card ID, the question asked, and the substantive answer, 2-4 sentences each]

For each card: read it, append insights under ## Discussion Notes per the update protocol, set date_updated, write the card back.

Anti-Patterns

• NEVER write more than 5 sentences per paper. This is a briefing, not a deep dive. If you're writing multi-paragraph walkthroughs, you're doing cross-paper-report's job.
• NEVER mirror the paper's self-assessment of novelty. Every paper claims to be revolutionary. Your job is to say plainly what it actually does.
• NEVER skip the Recommendation section. The whole point is to help the user decide where to focus. Be opinionated.
• NEVER generate a briefing with zero source material. If no KB cards, conversation context, or paper search results are available, ask the user for a topic or paper URLs.

Graceful Degradation

• No KB configured: Use Path B (conversation context) or Path C (paper search). Note in the output that a KB-sourced briefing would be deeper.
• Few cards (1-2): Still produce the briefing, but note that coverage is thin and suggest adding more papers via $paper-discovery or $paper-synthesis.
• Cards lack depth: If cards are shallow (abstract-only synthesis), note this in the briefing and suggest re-running $paper-synthesis with full PDF access.
• Path C paper search returns few results: Broaden the search facets or note that the topic is niche. A thin briefing is better than no briefing.