Method
How a hunt works
An agent searches public sources for a problem people pay to solve, checks who already sells a fix, and writes a verdict with every source linked. This page says how, step by step, including where it goes wrong.
1.0 The path
Six phases, from a sentence to a repository
The app shows which phase each hunt is in. The agent runs the first four, and you decide at the points where a person should: whether to keep a candidate, and whether to go on.
- LookingWrites searches from your instruction, runs them and files each finding with its link.Agent
- SortingGroups findings into markets. A market that pays must name a priced example.Agent
- CheckingWho sells it and at what price, what their users complain about, four ways to win.Agent
- ShapingWhat you would build at four sizes, each graded on reach, edge and effort.Agent
- BuildingYou take a size to a project: files, a thesis and a stop condition.You
- Handed offA build brief, downloaded or pushed to your own GitHub.You
2.0 Sources
Where it looks
Public pages only. How thorough a hunt is, Quick, Standard or Deep, is your choice, and the start screen says what each costs.
- The public web
- Through search, for products, prices, reviews and complaints.
- App stores
- The App Store, WordPress.org plugins and Firefox add-ons through their own public interfaces; Google Play and the Chrome Web Store as pages. Reviews of three stars or fewer come from seven stores in a structured way, and from about a dozen more as page text.
- Request boards and issue trackers
- Of products that already sell, and the issues of open source products, where users say what is missing.
- A repository index
- Open source repositories ranked by meaning and by words together, for engines that do the hard part but have no interface a non-programmer can use.
- Forums and communities
- Where people complain about a product or ask for one.
- The sparks feed
- A separate daily read of Reddit threads, freelance postings, low reviews of top-grossing iPhone apps, Stack Exchange recommendation questions and Hacker News. Each need is kept as a paraphrase, a short quote and a link, never a username.
3.0 Evidence
How evidence is kept
Nothing is cited that is not on file, and nothing a web page says can steer the agent.
- Every page is saved
- Each page the agent reads goes to your library with the search that fetched it.
- Citations are numbered
- A summary may cite only the sources it was given, by number, and name only products the market check found. Anything else is dropped before you see it.
- Links must come from a source
- A link in a result must be one a source returned. A document refuses a link that is not already in your library, because an address built during a run is what a hostile page would try to get written down.
- Page text is fenced
- Text from a web page, an upload or a draft is marked untrusted when the agent reads it. Once a turn has read such text, it cannot follow a link that the text suggested.
- What is shared, and what is not
- Public pages are cached and shared between accounts to save cost. What you searched for, and what you wrote about them, is not.
4.0 Verdicts
How a verdict is reached
Each candidate gets one summary in three parts, each graded from 0 to 10, and one of three verdicts.
- The problem
- The people on file who asked for something: what they do today, what they spend, and the link to where they said it.
- The idea
- What would be built, and for whom.
- Can you win
- The products that already sell, their prices read from their own sites, and the gaps their users report.
Worth buildingWorth a closer lookNot worth building
The rules a verdict is held to
- Fewer than two sources
- If fewer than two linked sources show the need, worth building is held at worth a closer look. It is never turned into a no, and it never stops you keeping the idea.
- Nobody paying
- If nothing on file shows anyone paying, the verdict is held the same way.
- One verdict, said once
- The agent reports the verdict the card shows. It does not give a second opinion in chat unless it says it disagrees.
- A product site that does not load
- The table says so and does not borrow another page.
See a real one, a no, in the sample verdict.
5.0 Invented ideas
The one step allowed to invent
One step, go further, may propose an idea by borrowing a principle from a distant field. Each idea must answer a need the hunt found, and it is checked against the web and the repository index for something that already does it.
It is stored as invented, badged as invented on screen, and never counted as evidence that a market exists. Evidence is what people said and what products exist, never what the model thought of.
6.0 Cost
What it costs
Work is priced in credits. New accounts get 200 free, with no card. Credits never expire and nothing renews.
- HeldBefore a run starts, its maximum price is set aside from your balance.
- ChargedWhen it ends, only what it used is taken.
- ReturnedA run that fails or is cancelled returns everything.
| Typical action | Credits |
|---|---|
| A first research turn | about 40 |
| A chat message | 2 |
| A build brief | 10 |
| A pack of 700 | $10 |
| A pack of 2,000 | $25 |
| A pack of 5,000 | $50 |
Every charge is on your billing page. Prices include VAT where it applies.
7.0 Limits
Where it goes wrong
Every claim has a link so that you can check the ones that matter. These are the ways it fails.
- Models misread pages
- A price taken from the wrong tier, a product matched to a namesake, a complaint read as stronger than it is. Open the link.
- Public pages are not the whole market
- Private sales, enterprise contracts and word of mouth leave no trace. An empty search is not proof that nobody is buying.
- Some stores cannot be read
- A few catalogues block readers or have no public reviews. They are logged as not checked, never reported as an empty market.
- Search can fail in one place
- If the web search is unavailable, the check says it could not run. It never calls an idea open without the web.
- Grades are judgement inside rules
- For the four sizes, edge, effort and net value are counted from the plan by formula. Reach is the model’s estimate, and it carries the figure it rests on. All of them depend on facts a model read.
- It does not forecast
- Worth building means the evidence points that way today. It does not mean customers will arrive.
- It does not copy
- The rebuild tools describe what a product does from its public pages. They never download, decompile or run it. What you build, and the rights in it, are your responsibility.
8.0 Your data
Where your work goes
- Isolated per account
- Each account’s data is separated from every other account’s at the database level, on servers in the European Union.
- Sent only to do the work
- Your messages, the pages being read and any files you upload go to AI and search providers through paid API services. We have not opted in to any training on that content. The privacy policy and the subprocessor list name every company involved.
- GitHub
- A project can push its build brief to a new repository in your own account, with a README and guide files for the tool that writes the code. The repository is public, and the page says so before you agree. The app never pushes workflow files.
- Your own AI tools
- A read-only connection lets tools such as Claude or Cursor read your projects, briefs, hunts and candidates. It cannot start a hunt, change a file, spend credits or push anything, it allows 60 calls a minute, and every call is logged.
- Closing your account
- Deletes your content.
Questions
Where does Hunting Apps get its information?
From public pages only: web search results, app store listings and their low-rated reviews, request boards and issue trackers of products that already sell, forum complaints, freelance postings and a search index of open source repositories. Every page it reads is saved and linked from the result, so you can open the same source it used.
Can it tell me an idea will make money?
No. It reports what it found: who already sells something similar and at what price, who is asking for a fix, and where existing products leave a gap. A verdict of worth building means the evidence on file points that way. It is not a forecast, and a thin market can still hide a customer the public web does not show.
What stops it from inventing sources or competitors?
A summary may only cite sources it was given, numbered one by one, and may only name products the market check found. A link in a result must be one a source returned, and a document refuses links that are not in your library. Text the agent reads from a web page is treated as untrusted and cannot tell it where to go next.
What does a check cost, and what if it fails?
You pay in credits. A first research turn is typically around 40 credits and a chat message 2. Before a run, its maximum price is set aside, afterwards only what it used is charged, and a run that fails or is cancelled returns everything. The billing page shows every charge.
Does it see my ideas, and is anyone else able to?
Your ideas, files and projects belong to your account and are visible to no other account. Work is sent to AI and search providers only to do what you asked. The privacy policy and the list of providers say exactly what goes where, and closing your account deletes your content.
Bring the idea you can’t stop thinking about.
200 free credits. No card.