- 5,621 signals
- 10 sources
- 60 published ideas
- 115 payment-verified
Kill weak startup ideas
before your AI builds them
Describe your idea. CodeKudo attacks it with real complaints, spending signals, competitors and the reasons it fails — then tells you to build it, test it, narrow it, or stop. If it survives, you get the spec in the exact format your coding tool expects.
One spec, 30 correct formats
CLAUDE.md + AGENTS.mdCursor.cursor/rules/*.mdcWindsurf.windsurf/rules/*.mdGitHub Copilot.github/copilot-instructions.mdOpenAI Codex CLIAGENTS.mdGemini CLIGEMINI.md + AGENTS.mdAWS Kiro.kiro/specs/{slug}/requirements.md · design.md · tasks.mdGoogle AntigravityAGENTS.mdZedAGENTS.mdCline / Roo Code.clinerules + AGENTS.mdDevinAGENTS.md + knowledgeAmpAGENTS.mdTraeAGENTS.mdAiderCONVENTIONS.mdLovableKnowledge Base + founding prompt6-part founding promptBase44Single full-stack promptEmergent.shAgent-segmented brief (UI / logic / DB / API / QA)Replit Agentreplit.md + numbered task listv0 (Vercel)Per-screen component prompts + design tokensSoftgenFounding prompt + iterationsCreate.xyzFounding promptDatabuttonPrompt + Python backend briefTempoFounding prompt (React)TrickleFounding promptVybeFounding promptOnSpace.AIFounding promptRorkReact Native + Expo prompt + screen flowFireVibeScreen list + brand tokens + handoff bridgeFlutterFlow AIVisual + AI brief (Flutter)Three claims, and each one is checkable
Supporting
- Store operators describe returns as manual work costing roughly 11 minutes per order.
- Twelve people said outright they would pay for this rather than keep doing it by hand.
Falsifying
- Shopify can ship this natively; they have shipped adjacent workflow features twice in 18 months.
- Two products serving this exact segment shut down, both citing support load per dollar.
Cause of death
Shopify adds native return rules and this becomes a feature, not a product. Your defensibility has to be the sub-$50k GMV segment the incumbents refuse to support — not the workflow itself.
We look for reasons to say no
Every card carries a falsifying column and a stated cause of death — the column the rest of this category leaves out.
×15
What money outweighs
Revenue confirmed with the payment processor counts 15 times a complaint. The weights are in code, and the verdict engine uses the same ones.
“Store operators describe returns as manual work costing roughly 11 minutes per order.”
Every claim resolves to a source
Click any reference and you get the platform, the date, a short excerpt and a link to the original. No record, no sentence.
- Claude Code
CLAUDE.md + AGENTS.md - Cursor
.cursor/rules/*.mdc - Windsurf
.windsurf/rules/*.md - GitHub Copilot
.github/copilot-instructions.md
Specs, not a PDF nobody reads
Pick your tool and get exactly what it expects — the file it reads, in the format it reads, under the limit it can digest.
This is what a verdict looks like
Evidence on the left, the case against it on the right, and a stated cause of death. The card below is the real component, filled with an example idea — every live card is built the same way from collected signals.
Example card — illustrates the format. Live cards are built from collected signals and every reference resolves to a real source.
Return automation for sub-$50k GMV Shopify stores
Rules-driven return approvals and label generation for small stores that the incumbents price out.
Demand ladder
Complaints are cheap. Money is the signal.
Counted from clustered complaint signals. No candidate-relative commercial check was applied, so no revenue is attributed to this idea.
Verified revenue: not established for this idea. No record ties a revenue figure to a product selling what this would sell.
3 revenue observations exist in this space, but none is tied to a product selling this outcome — so they are shown as context, not counted as evidence for this idea.
Confidence
Where the data is thin, we say so.
- Demand signal
- High
- Willingness to pay
- Low
- Market size
- No data
- Competitor gap
- Medium
47 signals across 6 sources
Only 3 direct signals — not enough to decide on
No direct data. Anything here would be a guess.
2 sources confirmed
Supporting evidence3
Store operators describe returns as manual work costing roughly 11 minutes per order.
Twelve people said outright they would pay for this rather than keep doing it by hand.
Three products in this niche have payment-verified revenue, the largest at $47k MRR.
Falsifying evidence3
Shopify can ship this natively; they have shipped adjacent workflow features twice in 18 months.
Two products serving this exact segment shut down, both citing support load per dollar.
Roughly 70% of the complaints are resolved with a free spreadsheet template people already share.
Most likely cause of death
Shopify adds native return rules and this becomes a feature, not a product. Your defensibility has to be the sub-$50k GMV segment the incumbents refuse to support — not the workflow itself.
Attack. Verdict. Spec.
Four answers, and one of them is no
A score out of 100 lets you round in your own favour. These do not. The verdict comes from published thresholds applied by code, not from a model deciding how it feels about your idea — so you can argue with the rule instead of trusting the tone.
Build candidate
The evidence supports building. This is not a promise that it will work — it is the absence of a reason not to start.
Validate first
Something is here, but not enough to commit. Spend a week testing the gap before writing code.
Narrow the wedge
The idea as stated is too broad for the evidence. A smaller version of it is defensible.
Kill it
The evidence contains a reason not to build this. Read it before you argue with it.
A build candidate needs at least 8 signals across 2 platforms, one signal where money actually moved, and no unmitigated fatal risk. Missing any one of those downgrades the verdict rather than lowering the bar — every threshold is published.
A complaint is not a customer
Idea tools count mentions. Mentions are free to produce and cost nothing to make, which is why a thread full of people agreeing with you feels like a market and usually is not.
Every signal here is typed by what it proves, and the four types are not worth the same. The weights are fixed in code and the same ones the verdict engine uses.
- ×1Somebody complained“This workflow is a nightmare.” Free to say, and most of the internet is this.
- ×3Somebody said they would pay“I'd pay for a tool that did this.” Stated intent, and stated intent is cheap.
- ×8Somebody is already spendingA paid tool named, a contractor hired, a budget line quoted.
- ×15Somebody is already being paidRevenue confirmed with the payment processor for a product in this space.
The chain, unbroken
The tools in this category each own one link and hand you the gap. A complaint corpus with no answer. A perfect PRD with no data behind it. A 200-page report nobody acts on.
- 01
Signal
10 sources, pulled to the edge of every rate limit. Raw text is discarded; the derived signal and its link are kept.
- 02
Opportunity
Clustering, momentum over 30/90/365 days, and a demand ladder that weighs payment-verified revenue 15× a complaint.
- 03
Evidence
Complaint ↔ existing product ↔ revenue confirmed with the payment processor, plus the counter-evidence that argues against building it.
- 04
Decision
A verdict from published thresholds, answered from records only. No record, no claim — it will tell you the data doesn't support you.
- 05
Spec
One Universal Core, rendered into whatever your platform reads. EVIDENCE.md is generated by code, never by a model.
The three things nobody shows you
Each of these makes the product look worse in a screenshot. All three are the difference between research and a sales page.
What we say out loud
- How old the evidence isEvery signal carries its post date, and past 180 days the idea says so itself instead of reading as current.
- How many people saw itViews, specs generated and members building it — labelled as CodeKudo member activity, not market activity.
- What we could not checkEvery verdict ends with an inventory of missing evidence. It appears on positive verdicts too, where it is least welcome and most useful.
Where it comes from
- App Store
- Discourse forums
- GitHub
- GitLab
- Hacker News
- Product Hunt
- Stack Exchange
- EU procurement
- TrustMRR
Counted from the ingest log: sources that ran in the last 48 hours. We publish the derived signal, a short attributed excerpt and a link to the original — never a copy of the full text.
Two ways to run it
Type an idea and read the verdict yourself, or let the coding agent you already use ask on your behalf.
- Searching the live web
- Reading the pages that matter
- Weighing both sides
Type an idea, read the verdict
The whole run happens on this page. No account, and the run keeps going if you close the tab.
$ claude mcp add --transport http codekudo https://codekudo.com/api/mcp …
Connected · 9 tools available
codekudo-<idea>-claude-code.zip
├── START-HERE.md
├── CLAUDE.md
├── AGENTS.md
├── README.md
├── docs/
│ ├── PRD.md
│ ├── ARCHITECTURE.md
│ ├── ROADMAP.md
│ ├── GTM.md
│ └── EVIDENCE.md ← generated by code, never by a model
└── .codekudo/
├── product.json
└── manifest.jsonOr let your agent ask
9 MCP tools, 2 of them spend credits. The rest are free to call.
One idea, 30 correct formats
Pasting a 2,000-word PRD into Bolt drowns it. Bolt wants six sections under 400 words. Cursor wants a rules directory, not the deprecated single file. We ship what each one reads.
Already wrote a spec? Grade it first.
Paste what you were about to hand your coding agent. You get a score against the twelve things that make AI-generated code go wrong — missing scope boundaries, no data model, no acceptance criteria, no explicit non-goals.
Free, no account, and the text never leaves your browser — the grader runs client-side, so there is no request for us to log.
What it checks for
- Scope boundaries
- Explicit non-goals
- Data model
- Acceptance criteria
- Error and edge cases
- Auth and permissions
- External dependencies
- Success metrics
What is actually in here
Counted from the database when this page was built, not typed into the design. The catalogue is small on purpose: an idea is published when its evidence holds up, which is a slower process than generating one.
5,621
typed signals
extracted from 103,580 records across 10 platforms
116
problem clusters with real weight
five or more independent signals; thinner ones exist but are not presented as opportunities
60
published ideas
30 readable in full without an account, evidence chain included
115
products with payment-verified revenue
confirmed with the payment processor through TrustMRR — the only tier that counts as proof somebody pays
30
coding tools with a correct format
the file each one reads, in the format it reads — not one PRD pasted everywhere
2h
since the newest signal landed
every connected source runs at least once a day
Sources: App Store · Discourse forums · GitHub · GitLab · Hacker News · Product Hunt · reddit · Stack Exchange · EU procurement · TrustMRR. We publish the derived signal, a short attributed excerpt and a link to the original — never a copy of the full text.
Where the catalogue is split
What the others left out
We went through every idea-database and idea-validator tool we could find and checked them against the same list. The category looks crowded from the side. Read down the column and it empties out.
| Fundamental | Other tools | CodeKudo |
|---|---|---|
| Real signal data | Standard | |
| Clickable source for every claim | Standard | |
| Payment-verified revenue | Some, partially | |
| Platform-specific spec output | Some, partially | |
| Evidence traced into the spec | Some, partially | |
| First-100-users plan from the data | Some, partially | |
| Founder-fit score | Some, partially | |
| Outcome feedback loop | Some, partially | |
| Counter-evidence engineonly here | Nobody | |
| Weighted payment evidenceonly here | Nobody | |
| Freshness and momentumonly here | Nobody | |
| Saturation transparencyonly here | Nobody | |
| Live monitoring and alertsonly here | Nobody |
Five tools surveyed. “Some, partially” means at least one does a version of it, not that the category does. Anything markedonly here was found in none of them.
Bring the idea you can't stop thinking about
We'll show you the strongest case against it. If it still stands, you leave with a spec your coding agent can read.
Attacking an idea and grading a spec are free and need no account · what membership costs