---
name: deep-research
description: >
  Multi-step research orchestration in the spirit of Claude/Gemini deep research.
  Use when the user wants a thorough investigation, comparison, "find everything about…",
  "research…", a well-sourced answer with multiple references — not a quick one-shot search.
---

# Deep Research

You orchestrate a multi-step investigation: decompose the question into
sub-questions, gather evidence from multiple sources, cross-check findings, and
synthesize a structured report with citations. Optimize for depth and
verifiability, not speed.

## Procedure

### 1. Plan (always first, visible to the user)

- Restate the question in one sentence to confirm scope.
- Decompose into **3–6 sub-questions** covering different axes of the topic.
- Show the plan briefly ("I'll split this into: …") and proceed — do not wait for
  approval unless the request is genuinely ambiguous.

### 2. Gather

For each sub-question in sequence:

1. `web_search` — find relevant sources (DuckDuckGo, up to 8 results).
2. `web_fetch` on the 2–4 most promising results — read the actual page content,
   not just the snippet.
3. Note: concise findings + the **URL of every source used** + confidence level.

### 3. Verify

- Cross-check findings across sub-questions and sources.
- **Flag contradictions explicitly** ("source A claims X, source B claims Y") —
  do not paper over them.
- Mark claims supported by a single source as unverified.
- If an axis of the topic is under-covered, run one more `web_search` + `web_fetch`
  round before synthesizing.

### 4. Synthesize (report)

Output structure:

```
## Shrnutí
2–4 sentences directly answering the original question.

## Zjištění
Organized by sub-question / axis. Every non-trivial claim carries an [n] citation.

## Rozpory a nejistoty
Where sources disagree, what could not be verified. (Omit the section if none.)

## Zdroje
[1] Title — URL
[2] …
```

## Progress reporting

Deep research can take several minutes. **Emit a short status message between
phases** so the user (especially on Telegram, where there is no thinking stream)
sees the task is alive. Examples:

- After step 1: `Plán: 5 podotázek — (1) … (2) … (3) …`
- After each sub-question: `[2/5] kimi-k2 benchmarks — 3 zdroje, hotovo`
- Before step 4: `Všechny podotázky pokryty, syntetizuji report.`

Keep status lines to one short sentence. No filler, no emojis. The final report
comes as a separate, full message at the end.

If a `web_fetch` fails or stalls, say so in a status line and continue — do not
abort the whole run silently.

## Rules

- **Always cite URLs.** Claims without a source must be labeled as your own
  inference / estimate.
- Prefer primary and recent sources; for fast-moving topics, watch publication dates.
- Do not invent facts. If something cannot be found, say "not found" — do not guess.
- Length proportional to the question. No filler.
- **Respond in the user's language.**

## Tools used

`web_search` · `web_fetch` · `write_file` (optional: persist the report under `workspace/`).