npx skills add senavoh/wedge
Then ask your agent:
Use wedge to find me a business I can start with $500 and 10 hours a week.
Works with Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and any agent that reads SKILL.md files.
Why Wedge
Ask an AI for business ideas and you get confident, generic, unsourced lists. Wedge works differently:
- Evidence or it didn't happen. Every pain point and claim about a competitor links to a real review, thread or pricing page.
verify.py re-fetches each source and checks that the quote is actually there. A claim needs two verified sources from different URLs, or it's labeled unverified and can't drive the verdict.
- It argues with itself. A separate critic pass tries to kill the top idea (who already tried, a free incumbent, people who complain but won't pay) before anything is ranked.
- It's honest. The verdict is GO, PIVOT or KILL, with the reasoning shown. If it couldn't reach any real customers' words (every source blocked or empty), it says INCONCLUSIVE and tells you how to finish, instead of guessing.
- It tells you how to win, not only what to build. It finds switching triggers, positioning in customers' own words, offers that pull switchers, where unhappy customers hang out, and drafts comparison pages and outreach.
- It's free to run. No API keys, no paid data. The scripts use only the Python standard library.
- It picks up where it left off. Everything is saved in
ideas/<slug>/, so a run can continue tomorrow or in a different agent.
Four ways to start
| Mode | Say something like |
|---|
| From scratch | "Find me a business I can start." |
| From what you know | "I run an IT company. What could I build?" |
| Stress-test an idea | "Is a dog-walking app for apartment buildings any good?" |
| Beat my competitors | "How do I take customers from Jobber?" |
Choose quick (about 5 to 10 minutes) or deep (the full run; our timed deep run took 29 minutes and read over 1,400 pages).
How it works
founder fit → niche scout → competitor map → review mining → gap ranking → critic → win plan → launch kit → verify → report
- Founder fit. Five quick questions (budget, skills, hours, risk, online or local) so ideas are ranked for you.
- Niche scout. Mines "I wish there was…", "alternative to…" and "too expensive" posts to find unmet demand.
- Competitor map. Real pricing, positioning and a feature matrix, all cited.
- Review mining. Turns hundreds of 1 to 3 star reviews and complaint threads into ranked pain themes with exact quotes.
- Gap ranking. Scores six axes (pain, willingness to pay, competition gap, founder fit, reachability, build cost) and shows the reasoning.
- Critic. A skeptical second pass tries to kill the winner.
- Win plan. Switching triggers, positioning, offers, what to build first, where to find unhappy customers, and draft assets.
- Launch kit. A name with a verified-available domain, pricing, the cheapest stack, a landing page, and a 7-day test with a pass bar and kill criteria.
- Verify and report. Every quote is checked, then rendered into one self-contained HTML report and a share card.
See real runs
16 real runs. 0 GO. That isn't a bug. Wedge only says GO when the pain is sharp, people already pay, nobody serves them well, you can reach them for free, and a skeptical critic can't find a sourced reason it fails. Most markets with public complaints already have someone serving them, so most ideas come back PIVOT or KILL. Better to hear that before you build than after.
Every run below is unedited, and every quote links to its source. Re-check any of them with verify.py. One of them is Wedge stress-testing a paid version of itself (KILL), and one is Wedge testing its own pivot (the score went down).
What would a GO take? See what each verdict means. If you get a real one, open a pull request and add it under examples/.
The skills
| Skill | What it does |
|---|
wedge | Runs the whole pipeline, keeps state, verifies and renders the report. |
wedge-niche-scout | Finds niches with unmet demand that fit you. |
wedge-competitor-map | Pricing, positioning and a feature matrix for every competitor. |
wedge-review-miner | Pain themes from real reviews and threads, with exact quotes. |
wedge-gap-ranker | Scoring, the critic pass, and GO / PIVOT / KILL. |
wedge-win-plan | How to pull customers from competitors. |
wedge-launch-kit | Name and domain, pricing, stack, landing page, 7-day test. |
wedge-watch | Monthly re-check of competitors: new complaints, price changes, new entrants. |
Each skill works on its own too: "Use wedge-review-miner on Jobber and Housecall Pro."
Install
Any agent (recommended)
npx skills add senavoh/wedge
The skills installer needs Node.js 22 or newer. On older Node versions, use the install script below.
Claude Code plugin
/plugin marketplace add senavoh/wedge
/plugin install wedge@wedge
Install script. Copies the skills into every agent skills folder it finds (~/.claude, ~/.codex, ~/.gemini, ~/.cursor, ~/.config/opencode, ~/.agents):
curl -fsSL https://getwedge.dev/install.sh | sh
To choose the folders yourself, pass them after sh -s --:
curl -fsSL https://getwedge.dev/install.sh | sh -s -- ~/.claude/skills ~/.codex/skills
Manual. Copy the folders in skills/ into your agent's skills folder.
Requirements: Python 3.8 or newer, which is preinstalled on macOS and most Linux systems.
Optional, for Reddit. Reddit blocks anonymous requests from many networks. Without keys, Wedge uses your agent's own web search for Reddit (site:reddit.com) and leans on sources that need no keys, such as app store and plugin marketplace reviews. If you already have approved Reddit API access, set REDDIT_CLIENT_ID and REDDIT_CLIENT_SECRET to use the official API. New Reddit apps now need Reddit's approval under its Responsible Builder Policy, so don't count on getting keys quickly. GITHUB_TOKEN, if set, raises GitHub's rate limit.
The scripts
The agent runs these for you. They live in skills/wedge/scripts/ and you can run them yourself too:
| Script | Does |
|---|
fetch.py | Pulls customer voice from Hacker News, Reddit, App Store reviews, GitHub issues and any web page, as JSONL. pages fetches many URLs in parallel, and links collects review links to feed into it. |
signals.py | Turns raw records into a short, cited digest of pain and demand signals, so the agent reads 100 lines instead of 100,000. |
verify.py | Re-fetches every cited source, checks the quote is there, and applies the two-source rule. |
domain.py | Checks domain availability through public RDAP, and reports "unknown" instead of guessing. |
state.py | Saves progress so any agent can resume a run. |
report.py | Renders the HTML report and the share card. |
How much does a run cost?
Wedge uses your agent's normal usage. There are no API keys and no paid data. It's built to keep that usage low:
- Quick mode is the default. Deep mode runs only when you ask for it.
- Raw pages never enter the conversation. The scripts save them to disk, and the agent reads a short digest instead.
signals.py --find pulls out only the sentences worth quoting.
- Extras are on request. In quick mode, the landing page and the outreach and comparison drafts are offered at the end instead of always built.
Measured across several test runs with the same model, a quick run used about 100k tokens and took 2 to 6 minutes. A run that hits the evidence floor (150+ records, 15+ customer quotes) costs about the same, because the extra pages are saved to disk, not read. A deep run, which scouts 7 niches and reads over 1,400 pages, took 29 minutes and used a few times that. To spend even less, run a single skill, for example "use wedge-review-miner on Jobber and Housecall Pro".
Data and ethics
- Wedge reads public posts and reviews through official or public endpoints: the Hacker News Algolia API, Reddit's JSON and official API, Apple's public review feeds and the GitHub API. It doesn't scrape sites whose terms forbid it. G2, Capterra and Trustpilot content comes in only through ordinary web search results.
- Quotes are short, attributed and linked, used to support analysis.
- The win plan never suggests fake reviews, astroturfing or misleading comparisons.
Contributing
Issues and pull requests are welcome. Good first contributions: new data sources that need no keys, better signal patterns for non-English markets, and example runs in new industries (add them under examples/).
License
MIT
