71 lines
3.1 KiB
Markdown
71 lines
3.1 KiB
Markdown
|
---
|
||
|
|
name: ask-research
|
||
|
|
description: "Dispatch web research to the research profile."
|
||
|
|
version: 1.0.0
|
||
|
|
author: Hermes Agent
|
||
|
|
license: MIT
|
||
|
|
platforms: [linux]
|
||
|
|
metadata:
|
||
|
|
hermes:
|
||
|
|
tags: [research, delegation, web-search, scraping]
|
||
|
|
related_skills: [ask-hermes, deep-research, better-search]
|
||
|
|
---
|
||
|
|
|
||
|
|
# ask-research — Generic Research Profile Dispatch
|
||
|
|
|
||
|
|
## Overview
|
||
|
|
|
||
|
|
Dispatches a research task to the research profile (`hermes -p research`). The research profile has a full arsenal of web research tools — SearXNG, web scraping, browser automation, social media scraping, Python analysis — and runs with `reasoning_effort: high`. No methodology skill loaded by default — the research agent uses its own judgment to pick the right tools for the question.
|
||
|
|
|
||
|
|
This is the generic "go research this" dispatch. For exhaustive methodology-driven research, use `deep-research` instead (loads the `deep-web-research` skill with six-move flow).
|
||
|
|
|
||
|
|
## When to Use
|
||
|
|
|
||
|
|
- "ask research to find everything about X"
|
||
|
|
- "research this topic and give me a detailed report"
|
||
|
|
- Any web research task that needs more than a few searches but doesn't need the full deep-research methodology
|
||
|
|
- Gathering documentation, community knowledge, official sources on a specific tool/technology
|
||
|
|
|
||
|
|
## When NOT to Use
|
||
|
|
|
||
|
|
- Simple factual lookup → use `web_search` directly
|
||
|
|
- Exhaustive OSINT/person investigation → use `deep-research` (methodology-driven)
|
||
|
|
- Task needs your exact context → use `delegate_task`
|
||
|
|
|
||
|
|
## Command
|
||
|
|
|
||
|
|
```bash
|
||
|
|
terminal(background=true, notify_on_complete=true, timeout=14400,
|
||
|
|
command="hermes -p research chat -q \"<prompt>\" -Q --max-turns 4000 --yolo",
|
||
|
|
workdir="/home/n8n/workspace/research")
|
||
|
|
```
|
||
|
|
|
||
|
|
- Always use `--max-turns 4000` (operator standing rule: 4000 is the safety net, not a budget)
|
||
|
|
- Always use `background=true` + `notify_on_complete=true` — research runs long
|
||
|
|
- The research agent writes output to the path specified in the prompt
|
||
|
|
- Capture the session_id from the process log after completion
|
||
|
|
|
||
|
|
## Prompt Composition
|
||
|
|
|
||
|
|
Every `ask-research` prompt must include:
|
||
|
|
|
||
|
|
1. **Exact research targets** — URLs, topics, sources to cover
|
||
|
|
2. **Output path** — absolute path for the deliverable file
|
||
|
|
3. **Web search mandate** — "Use mcp_searxng_searxng_web_search for EVERY claim. Scrape full pages with web_extract. Cite source URLs inline."
|
||
|
|
4. **Scope boundaries** — what to include AND what to exclude
|
||
|
|
5. **Turn budget note** — the agent has 2000 turns, use them thoroughly
|
||
|
|
|
||
|
|
## Post-Dispatch Behavior
|
||
|
|
|
||
|
|
Same as `deep-research`: return control to operator immediately. Do NOT poll. Do NOT check. Do NOT auto-deliver. Only check for results when the user explicitly asks.
|
||
|
|
|
||
|
|
## Spot-check Rule
|
||
|
|
|
||
|
|
Research agent self-reports are not verified fact. If it claims a file write, `read_file` the path to confirm. If it claims a specific finding, spot-check the source URL.
|
||
|
|
|
||
|
|
## Related Skills
|
||
|
|
|
||
|
|
- `deep-research` — Exhaustive methodology-driven research (six-move flow, external ledger, disconfirmation)
|
||
|
|
- `better-search` — Medium-depth research (1-3 loops with AI evaluation)
|
||
|
|
- `ask-hermes` — General-purpose peer agent dispatch
|