内容
# hand-raisers
A Claude Skill that finds the people who publicly asked for what you sell this week, qualifies each ask, checks how crowded the thread already is, and drafts the reply.
Cold outbound interrupts someone who never asked. This does the opposite: it finds the ones mid-question. "Anyone got recommendations for...", "which tool for...", "looking for a tool that..." get posted every day, and they go cold in about a week. The whole skill is built around that window.
## What it does
| Stage | Tool | Output |
|---|---|---|
| Sweep | Bright Data `discover` | Real Reddit post URLs for 4-6 ask angles in your niche, ranked on a stated buying intent |
| Filter | Claude | Only titles that are a genuine question from a buyer about your category. Career advice, listicles and vendor posts are dropped |
| Read | Bright Data `web_data_reddit_posts` | Author, exact post date, full text, comment count, and every existing comment with its own date |
| Freshness gate | Claude | Anything older than 7 days is dropped, on the real post date, and the drop count is reported |
| Qualify | Claude | Fit (yes/partial/no), age, and how many real answers the asker already has |
| Draft | Claude | One reply per ask: answers the question, compares 2-3 options honestly, and places your offer as one of them |
Output is a table ranked by fit × freshness × low competition, top 3 first.
## Required connector
**Bright Data MCP.** The skill does not run without it: `discover` is what returns usable post URLs, and `web_data_reddit_posts` is what returns the post date and the existing comments. Both are needed, and there is no fallback that works.
Add it in **Settings → Connectors**, or for Claude Code:
```bash
claude mcp add --transport http bright-data "https://mcp.brightdata.com/mcp?token=YOUR_BRIGHTDATA_TOKEN&pro=1"
```
`pro=1` is not optional here. `web_data_reddit_posts` sits in the Pro tool set, so the free/rapid endpoint loads a server that cannot complete step 3.
Two things worth knowing before you run it:
- **Do not swap `discover` for a search engine call.** `search_engine` with `site:reddit.com "<phrase>" after:<date>` looks like the obvious way to do this. It is not: the google engine returns encoded `/goto?url=CAES...` redirect links instead of post URLs, and ignores `after:` entirely. The bing engine returns the Bing interface markup. Both were verified dead on 2026-08-27, and the skill tells Claude to avoid them.
- **Scraping `reddit.com/r/<sub>/new` directly needs KYC.** Bright Data gates Reddit behind a completed KYC form for residential access. Without it you get `Residential Failed (bad_endpoint)`. The skill treats this as an optional accelerator, says so when it is blocked, and does not retry in a loop.
## Install
**Personal (available everywhere):**
```bash
git clone https://github.com/ToolMonsters/hand-raisers ~/.claude/skills/hand-raisers
```
**Per project:** clone into `.claude/skills/hand-raisers` inside the repo instead.
**Claude apps (claude.ai / desktop):** upload `SKILL.md` in **Settings → Capabilities → Skills**.
Restart Claude, or start a new session, so the skill loads.
## Usage
```
/hand-raisers
find people asking for a cold email tool
who asked for CRM recommendations this week
```
Claude asks two things before sweeping: what you sell in one sentence, and 3-5 words your buyers would use in their ask. It derives the second from the first if you skip it.
## Built-in guardrails
- **It never posts.** No reply is sent on your behalf. You copy, edit, and post it yourself, from your own account.
- **No contact scraping.** If an asker left an email or a handle in their post, it is not collected or stored. You get the link to the thread.
- **Bad fits get skipped, not stretched.** A product forced into an ask it does not match costs more trust than staying quiet.
- **Age comes from the post, not the query.** Every ask is dated from `date_posted`, because the sweep step will happily return something from last year.
- **The funnel is reported.** How many results swept, how many survived the filter, how many dropped as stale. A high drop count means the queries were too generic, and you should see that rather than a short table with no explanation.
## License
MIT
---
Made by Yonathan Cohen (Tool Monsters)
MCP 市场