How to find a startup idea people actually want

Clawsmith's scanner currently tracks 1,503 demand signals built from 2,615 public posts across the OpenClaw ecosystem. Almost none of the strong ones started as someone's shower idea. They started as a complaint, posted in public, repeated by strangers who never met.

That is the whole method. People write down what they want every day. You can collect it, score it, and build the winner. This is the manual version, with the exact searches, a scoring sheet, and the agent handoff at the end. If you would rather have it run automatically, Clawsmith does this scan every 1 to 3 hours.

Step 1. Search where the complaints live

Three surfaces cover most of it. Budget 30 to 60 minutes per surface and save every hit into a spreadsheet: URL, exact quote, upvotes or reactions, date, username.

Reddit indexes poorly on its own search, so go through Google:

site:reddit.com "is there a tool that" claude code
site:reddit.com "i would pay for" ai agent
site:reddit.com/r/OpenClaw "i wish"

GitHub issues, in the biggest repos of your target space, sorted by thumbs-up reactions:

is:issue is:open sort:reactions-+1-desc label:enhancement

Sort by reactions, not comments. Comments are arguments. Reactions are lurkers quietly voting for the feature.

Hacker News through Algolia, filtered to comments from the past year:

hn.algolia.com/?query="does anyone know a tool"&type=comment&dateRange=pastYear

Swap the quoted phrases for your space. "Is there a tool that", "I would pay for", "why is there no", and "I ended up writing a script" are the four highest-yield patterns.

Step 2. Keep only problems that repeat

The pass bar is repetition: the same problem from 10+ distinct people across 3+ platforms. One viral thread is entertainment, not a market.

A real example from our data. One signal in the tracker reads "Text Blaze proves the browser snippet market, and no dominant AI prompt reuse extension exists." Its sources sum to 700,300 in engagement and the gap status is underserved. That is the shape you want: a category people already pay for, missing its obvious next version.

The counterexample is also in the data. The top signal right now is the Claude Code source leak, at 26,142,255 engagement across 8 platforms and 12 source posts. Massive number, zero product gap. It is a news event. Repetition of a problem beats size of a moment, every time.

Step 3. Confirm money already moves

A problem people complain about but will not pay to fix is a hobby. Before you commit, find proof of payment in the category.

In the OpenClaw ecosystem, 173 verified startups do roughly $300k a month combined on TrustMRR, and the top tool clears $107k a month on its own. That is what a paying category looks like. Check TrustMRR for your space, read pricing pages of the closest existing tools, and search "[category] revenue" threads. The bar is low but real: at least two products charging money, or one doing $10k+ a month.

Step 4. Score before you commit

Put every surviving problem through the same sheet. Copy these columns:

A row that fails two columns dies. Killing ideas fast is the point of the sheet. Most weeks you keep one candidate out of ten, and that ratio is healthy.

Step 5. Turn the winner into a build brief

A spec your coding agent can actually build from contains five things: the problem with source URLs attached, the target user, the two or three features that remove the pain, the stack, and a competitor table with each one's weakness.

This is the step Clawsmith automates end to end. Get an API key at clwsmth.com, then connect your agent:

claude mcp add --transport http clawsmith https://www.clwsmth.com/api/mcp \
  --header "Authorization: Bearer clwsmth_YOUR_KEY"

Or in JSON config for Cursor and friends:

{
  "mcpServers": {
    "clawsmith": {
      "type": "http",
      "url": "https://www.clwsmth.com/api/mcp",
      "headers": { "Authorization": "Bearer clwsmth_YOUR_KEY" }
    }
  }
}

Then ask your agent, in plain words:

Show me the top signals right now.
Get me the product brief for [the one you picked].

The brief comes back with market context, engineering-level feature specs with stack choices, the source posts with engagement numbers, and the competitive table. The agent writes REQUIREMENTS.md from it and starts building. Two MCP calls total.

Step 6. Your first customers are in the source posts

The people who complained are the outreach list. Across our 1,503 signals the extractor has pulled 67,449 leads, the actual usernames behind the complaints, classified as builders, complainers, or users.

The manual version works the same way: you kept the username next to every quote in your spreadsheet in Step 1. Your first message references the exact post they wrote. That is a warm open, because you built the thing they asked for in public.

When this method is not worth it

Three honest exits. If the category has a dominant incumbent and has existed for years, the gap is usually distribution, and a better product will not save you. If you cannot find 10 distinct voices after a week of looking, it is a hobby. If nobody in the space pays for anything, keep it as a side project and spend your building time elsewhere.

Summary

Search Reddit, GitHub issues, and HN with the query patterns above. Keep problems repeated by 10+ people on 3+ platforms. Confirm at least two products charge money in the category. Score everything against the sheet and kill rows that fail twice. Write the winner into a spec with sources attached, or connect your agent to the Clawsmith MCP and pull the brief in two calls. The live preview idea shows what a finished brief looks like, and the full flow with leads is on pricing.