validator

Validator

Validate a project or product idea before you build it. An agent skill that researches market demand, audience, competitors, moat, and gaps, and returns a GO / CONDITIONAL GO / NO GO verdict backed by sourced evidence.

What This Solves

43% of failed VC-backed startups cited poor product-market fit as a root cause of their death. That is 43% of the 385 companies whose reason for death CB Insights could identify (out of 431 analyzed) — not running out of money, not a bad team, not bad timing. They built something people did not want. (CB Insights, 2026)

It gets worse:

• Ran out of capital is the symptom, not the cause. It appears in 70% of failures, but CB Insights calls it "almost always the final cause of death, not the root problem." (CB Insights, 2026)
• The 431 failed companies raised a combined $17.5 billion before dying. The median company raised $11 million. They had money. They did not have evidence anyone wanted what they were building. (CB Insights, 2026)
• 81% of founders pivoted away from their original idea, and 42% wish they had pivoted or changed their business model sooner. (Wilbur Labs, 200 US tech founders, Wakefield Research, 2026)
• 54% of founders said the most important lesson they learned from failure was the need to better understand product-market fit. (Wilbur Labs, 2026)
• 74% of high-growth internet startup failures were attributed to premature scaling — growing teams, spend, and distribution before confirming the market actually wanted what they built. (Startup Genome, 2011)
• Roughly 90% of startups fail overall is a widely cited figure for venture-scale startups (Startup Genome, 2011); U.S. BLS data shows about 50% of new businesses survive 5 years. The most common proximate cause among failed startups — poor product-market fit — appears in 42-43% of post-mortems. (CB Insights / Startup Genome composite)
• Investors fund AI stories, not validated ideas. Y Combinator's 2025 cohorts were ~90% AI companies (154/170 in Summer, 140/154 in Fall — Inc., 2026), with 92% of the Fall batch building AI into their core offering (CB Insights). "AI"-branded startups raise 15-50% more than comparable non-AI startups (DePaul JATIP, 2026). But the label is not proof of demand: non-AI startups that merely claim AI raise better yet exit worse (CSWIM 2026, 16,713 US startups), and the SEC/DOJ now prosecute fake AI claims (Nate, PGI Global). AI gets you in the room; evidence keeps you alive.

The real problem is that "validation" is usually skipped or faked. Founders ask their friends, trust their gut, or confuse excitement with demand. By the time the market answers with silence, the money is spent and the team is built. Most startups never run a single search for their competitors. Most never define who exactly they are building for. Most never check if the problem is actually painful, or just "2 in 100 times" annoying.

That is how you end up in the 90%.

And Y Combinator knows it. In 2025 they poured funding into ~90% AI companies per batch (154/170 Summer, 140/154 Fall — Inc., 2026) — the highest concentration in the accelerator's history — many on little more than the word "AI" in the pitch. That is not diligence; that is a portfolio of unvalidated bets. Y Combinator, and every VC playing this game: this skill is for you. Validator is free, open source, and takes minutes, not hours. Run it on the next batch before you write the check. AI gets you in the room; evidence keeps you alive.

And if you are a founder pitching that same story — run Validator before you pitch. Every other founder is pitching an AI story too; the ones who survive are those whose story rests on evidence.

Validator is different. It forces the agent to research before you build. It forces ONE specific audience — not "everyone" — because "all businesses" solves nothing for anyone. It checks if the problem actually exists, if it is actually painful, if the market is growing, if competitors already solved it, and if your solution is buildable and legal TODAY. Every claim is backed by a source URL. The output is a verdict: GO, CONDITIONAL GO, or NO GO — with the exact conditions and a written falsification test so "maybe" cannot survive forever.

And it is open source. Free and open source — no per-report fee, no subscription.

What We Need

You.

Validator is open source because founders and solo builders deserve the same quality of pre-build research that funded teams get from analysts — without paying analyst prices.

Here is how you help:

• Run it. Next time you have an idea, run this skill before you write code. Tell us if the verdict was useful.
• Question the evidence. If a phase made a weak assumption, open an issue. The skill gets sharper every time a bad assumption gets caught.
• Share your reports. Did it save you from building something useless? Tell us your story. That is the whole point of this skill.

That's it. That's the whole thing.

Quick Start

This is an Agent Skills-format skill: a folder with a SKILL.md file containing YAML frontmatter and instructions. It works in any tool that supports the Agent Skills format.

# Clone
git clone https://github.com/Sumama-Jameel/validator.git
cd validator

# Copy the skill folder into your agent's skills directory
# Examples:
#   ~/.claude/skills/validator
#   .claude/skills/validator
#   .agents/skills/validator
#   ~/.config/opencode/skills/validator
cp -r SKILL.md examples ~/.claude/skills/validator

Or install straight from GitHub with the open skills CLI:

npx skills add Sumama-Jameel/validator -g

Or install it as a Claude Code plugin marketplace:

/plugin marketplace add Sumama-Jameel/validator
/plugin install validator@sumama-validator

Then just tell your agent you want an idea validated:

"I have an idea for X. Validate it."

The skill triggers on its own when you propose a new project or ask whether an idea is worth building.

Features

• GO / CONDITIONAL GO / NO GO verdict. The skill ends with an explicit decision, not a wishy-washy "it depends." If the verdict is CONDITIONAL GO, it lists the build conditions and requires a re-run when those conditions change.

• Falsification test for every CONDITIONAL GO. "CONDITIONAL GO" is not complete until it states a specific, observable condition that would flip the verdict to NO GO — e.g. "NO-GO if a funded competitor launches a $0 version within 6 months." It is a written exit gate, not a vibe.

• One audience, forced. "All businesses", "everyone", and "small and medium companies" are rejected as answers. The skill pushes until you pick the ONE segment you would build for first — because N different buyers with N different budgets and pains means the product solves nothing for anyone.

• 12 phases of evidence-based research. Problem validation → market research → every single competitor → head-to-head feature comparison → destroying your own assumptions → gap analysis → moat evaluation → outcome risk → final decision → pivot → final report.

• Every competitor, not just the famous ones. Enterprise, mid-size, startup, open source, side project — "a competitor is a competitor." Funding, revenue, pricing, employees, business model, launch year, all collected.

• Self-assumption destruction. "Our X feature is unique" gets checked against the market. If it is not unique, the skill says so in writing. No rose-colored glasses.

• Brutal gap and moat analysis. Gaps become opportunities. Moats are rated objectively — "Weak because everyone has it" vs "Strong because nobody offers it." Architecture elegance is explicitly not a user-facing differentiator.

• Platform policy and technical feasibility are part of validation. Before any GO/CONDITIONAL GO, the skill checks: does the product depend on a platform API or ToS that could break or get you banned? Can the core mechanic actually be implemented right now? If unverified, it is at most CONDITIONAL GO — never a clean GO.

• Every claim carries a source URL. Research answers are written into structured markdown files with the exact source URLs on the line below the information. No "trust me", no vibes, no unverified assumptions — which are STRICTLY FORBIDDEN.

• Data source priority. Government → industry reports → company filings → academic → surveys → official docs → reviews → social media → blogs. When sources conflict, the higher-priority source wins.

• Evidence saturation, not endless browsing. Stop searching when results become repetitive, after 3 consecutive empty searches, or when no new high-value evidence appears. It knows when it is done.

• Full report in a single markdown file. MAIN.md with market demand metrics, competitor tiers, problem-already-solved analysis, moat table, verdict, build conditions, falsification test, and pivot recommendations — in one structured document.

Architecture

                    +----------------------------------------------+
                    |              VALIDATOR                        |
                    |                                              |
                    |  Idea (raw user request)                     |
                    |         |                                    |
                    |         v                                    |
                    |  +---------------+   +-------------------+   |
                    |  |  Phase 0      |   |  Phase 1          |   |
                    |  |  Prepare base |   |  Validate problem |   |
                    |  |  Idea_Valid-  |   |  + FORCE ONE      |   |
                    |  |  ation/ folder|   |    audience       |   |
                    |  +-------+-------+   +---------+---------+   |
                    |          |                     |             |
                    |          +----------+----------+             |
                    |                     v                        |
                    |  +--------------------+    +--------------+  |
                    |  |  Phase 2-8         |    |  Phase 9-11  |  |
                    |  |  Basic research    |    |  Decision     |  |
                    |  |  Competitors       |    |  Pivot        |  |
                    |  |  Feature compare   |    |  Final report |  |
                    |  |  Assumption kill   |    |  MAIN.md      |  |
                    |  |  Gaps | Moats      |    |  Verdict      |  |
                    |  |  Outcome risk      |    |  Falsification|  |
                    |  +--------------------+    +--------------+  |
                    |                                              |
                    |  Files: MAIN | market | audience | problem    |
                    |  statement | solution | competitors.md       |
                    |                                              |
                    |  Rule: NO unverified assumptions.             |
                    |  Every claim -> source URL, every phase      |
                    |  reads previous phase files first.            |
                    +----------------------------------------------+

The 12-stage pipeline: Prepare base → Validate problem (force audience) → Basic research → Find competitors → Feature comparison → Destroy assumptions → Gap analysis → Evaluate moat → Outcome risk → Final decision → Pivot → Final verdict & report.

Files produced:

Competitor Comparison

Validator is the open source skill for pre-build validation. It has no subscription and no per-report fee.

The gap is not data — it is that almost nobody runs the research before building. Every other approach either skips the evidence entirely or outsources it to a human who guesses. Validator makes the agent do the work, with sources, before you spend a dollar.

The paid AI co-founder platforms are Validator, closed and metered. Cofounder.im's Idea Validator returns a confidence score (Yes / Proceed with Caution / Reconsider) for $9.99-$99/month plus 50 credits per run; Fonda charges $19-$39/month, scoutr $5-$79/month, FoundrAI $49-$59/month, IdeaOne $12-$19, and ScribeAI/Launchcrew one-time fees ($39/$249). None force one audience, require source URLs, or ship a falsification test — Validator does all three, for free, in your own repo.

Data Integrity Model

Project Structure

validator/
├── SKILL.md                  # The skill instructions (12 phases)
├── examples/
│   ├── example-report.md     # A complete sample final report (CONDITIONAL GO)
│   └── example-report-no-go.md  # A sample NO-GO report — honest-stop path
├── .claude-plugin/
│   ├── plugin.json           # Claude Code plugin manifest
│   └── marketplace.json      # Claude Code plugin marketplace catalog
├── README.md
├── LICENSE                   # MIT
├── CHANGELOG.md
└── .gitignore

The skill is designed to be dropped into any Agent Skills host. It writes its output files into an Idea_Validation/ folder in your current working directory.

Test

This is a behavior skill, so "testing" means running it on real ideas:

# Verify the skill is installable from this repo (open skills CLI)
npx skills add Sumama-Jameel/validator --list

Then run it on a real idea and check the output:

• Does it force one specific audience?
• Does every claim carry a source URL?
• Does the verdict include a falsification test?
• Are the gaps genuinely brutal (not cheerleading)?

Also Check Out

IntentPreter - Interpret and clarify ambiguous, confusing, or emotionally-charged user requests before acting. The skill that makes sure the agent understood what you meant before it does anything.

Eliminator - Debug complex software bugs using evidence-based investigation instead of guessing fixes. When the bug is weird and you are stuck, eliminate the wrong answers.

License

Licensed under MIT License.

Copyright 2026 Validator Contributors.

Built to answer one question before you build: does anyone actually need this? No assumptions. No vibes. Just sources and a verdict.