---
name: research-brief
description: Use when the user asks for a research brief, deep dive, explainer, literature scan, or "find out everything about" a topic, company, technology, or question. Produces a sourced brief with a bottom-line answer, key findings, open debates, and a numbered source list, with every claim tied to a source opened this run.
metadata:
  id: research-brief
  version: "1.0.0"
  author: BotPlaza
---

# Research Brief

> Any question, answered with the evidence and every source.

## What it does

Turn a question into a brief the user can act on: the answer up top, the evidence below, and every claim traceable to a source.

## When to use

- "Research X", "give me a deep dive on X", "what's the state of X", "should we use X or Y", "brief me on X before Thursday".
- Not for live prices, sports scores, or a single fact lookup. Answer those directly.

## Requires

Web search and web access to open sources. No accounts needed.

## Schedule

On demand. Run it when the user asks; it can also be put on a schedule (for example a weekly brief on a tracked topic).

## Inputs

| Setting | Default | Notes |
|---|---|---|
| `question` | required | Restate it as one clear question before starting. If it's vague, pick the most useful reading, say which, and proceed. |
| `depth` | standard | `quick` (5–8 sources, about 1 page), `standard` (10–15 sources, 2–3 pages), `deep` (20+ sources, with a methods note). |
| `audience` | the user | E.g. "my CFO", "a non-technical friend". Sets vocabulary and what to explain. |
| `recency` | any | Or "last 12 months", "since 2024". |
| `source_types` | balanced | Toggle: news, academic papers, official/government data, company filings, expert blogs, forums. |
| `format` | brief | `brief` (prose + bullets), `table` (comparison grid), `memo` (one page, decision-oriented). |

## Steps

1. **Frame.** Write the question in one sentence and list 3 to 5 sub-questions that together answer it.
2. **Search wide.** Run several searches per sub-question with different phrasings. Include primary sources (papers, filings, official data, documentation), not just articles about them.
3. **Read, don't skim.** Open every source you will cite. Note the publish date, author or organization, and the specific claim you're using.
4. **Weigh sources.** Prefer primary over secondary, recent over old (unless history matters), and independent over interested parties. Flag when a source has a stake in the answer.
5. **Find the disagreement.** Look for where credible sources conflict. Report the conflict honestly; don't average it away.
6. **Answer.** Write a 2–3 sentence bottom line that directly answers the question, with your confidence (high / medium / low) and why.
7. **Write the brief** in the format below. Tie each claim to a numbered source like [3].
8. **Check.** Every number and quote matches its source. Every source is listed with a working URL and date. Nothing comes from memory alone; if you know something but can't source it, say "unsourced" or cut it.

## Output format

```markdown
# {Question}
*Research brief · {date} · {N} sources · Confidence: {high/medium/low}*

## Bottom line
{2–3 sentences answering the question.}

## Key findings
1. **{Finding}.** {Evidence, 1–3 sentences.} [1][4]

## Where sources disagree
- {Point of conflict, who says what.} [2] vs [6]

## What we still don't know
- {Gap}

## If you're deciding
- {Practical implication or next step}

## Sources
1. {Author/Org}, "{Title}", {Outlet}, {date}. {URL}
```

## Edge cases

- **Little reliable information:** say so in the bottom line. A short honest brief beats a long speculative one.
- **Fast-moving topic:** add "As of {date, time, tz}" and note what may change.
- **Medical, legal, or financial questions:** summarize the evidence, note it isn't professional advice, and point to the right kind of professional.
- **Paywalled papers:** use the abstract and say only the abstract was read.
- **The user's premise is wrong:** say so plainly in the bottom line, with the source.

## Boundaries

- Read-only. This skill never sends, posts, or buys anything.
- Never cite a source you didn't open this run, and never present model memory as a finding.