An AI venture analyst that runs in your terminal. It interviews you, researches the market with real sources, scores your idea across six dimensions, and hands you a verdict with a two-week experiment to test the one assumption most likely to kill it.

It is built to disagree with you.

Maintained by Latif Abderrahmane

See two real verdicts → · 60-second start ↓

If you have ever asked an AI to generate a million-dollar idea and believed the answer, star this.

The problem this exists for

Ask a general chatbot whether your idea is good and it will find a way to say yes. It has no rubric, no memory between steps, no obligation to cite anything, and no mechanism that lets a single fatal flaw outweigh five encouraging paragraphs.

FindMeSaaS is the opposite by construction:

Scoring is multiplicative, not additive. Any dimension below 25 out of 100 multiplies the whole score down. One catastrophic weakness cannot be averaged away by four good ones, because that is not how startups fail.
Every claim carries a URL. Competitor prices, market sizes and channel costs are written to disk with the page they came from and the date it was read.
It scores its own confidence. Missing a dimension discounts the score. It will tell you it does not know.
It has a test suite. A 700-line dependency-free harness re-computes every score, checks every output against a contract, and fails the run if the arithmetic does not hold.

60-second start

No API keys, no configuration, no dependencies.

Add it to any project

npx find-me-saas init

That is it. Open the folder in Claude Code, Codex or Cursor and the commands are there.

Useful flags:

npx find-me-saas init --dry-run              # show what it would write, change nothing
npx find-me-saas init --platform claude      # claude | codex | cursor | all (default all)
npx find-me-saas init --force                # overwrite existing tool files

It will never overwrite your analyses. memory/user_profile.md, memory/ideas/ and your research files are protected even with --force, and every protected path is printed. A reinstall cannot cost you work.

Or clone the repo

git clone https://github.com/Latifox/find-me-saas.git
cd find-me-saas
claude      # or: codex

Cursor users open the folder and the chat; rules load from .cursor/rules/.

Nine commands

Type / and they are all there. Each one does a single job.

Command	What it does	Time
`/founder-profile`	Build or refresh your profile: background, constraints, target buyer	15 sec – 5 min
`/find-idea`	Research a market, generate and rank 7–10 candidates matched to your edge	~15 min
`/validate-idea`	The full ten-step chain, ending in a decision memo	~15 min
`/gut-check`	Four dimensions, no memo. Can rule an idea out, never rules one in	~3 min
`/market-scan`	Trends, competitors, sizing, channels. Add `--quick` for trends only	5–15 min
`/pivot-idea`	Root-cause the weak dimensions and generate evidence-backed pivots	~10 min
`/auto-pilot`	Autonomous. Onboard, research, rank, fully validate the best one, hand you the memo	~25 min
`/idea-status`	Portfolio view of every idea, its score, and what is still missing	instant
`/verify-memory`	Run the test suite over your analyses and explain anything it flags	instant

Or ignore all of them and just talk. "validate my idea: ..." routes to the same place.

About `/auto-pilot`

It runs the entire pipeline without stopping to ask. Where it would normally ask a question it takes the documented default, records the assumption, and keeps going; gaps land in the memo's watermark rather than in your inbox. It stops early only for three things: no usable market signal, too few dimensions to score honestly, or a harness error it cannot fix. It costs real research time and produces a real analysis, so it tells you that before it starts.

Two hooks keep it honest

Installing for Claude Code registers two hooks, and both exist because a prompt system that only checks itself when asked eventually stops checking.

On session start, it reads your profile and idea directory and tells the agent where things stand: whether you have been onboarded, which constraints are still unset, what you have already analysed, and which ideas were started and abandoned. No more re-explaining yourself at the top of every session.

After any write into an idea directory, it runs that idea through the validation harness. If the file breaks its contract — a score whose arithmetic does not reconcile, a verdict that does not match the threshold table, a missing source array — the errors go straight back to the agent while it still has the context to fix them. The write is not blocked, because it already happened; the agent is simply told what it got wrong.

Both are plain Python with no dependencies, so they behave the same on Windows, macOS and Linux. They fail open: a broken hook exits quietly rather than breaking your session.

It starts by asking who you are

The first thing FindMeSaaS does is onboard you, because everything downstream depends on it. Founder-market fit is scored from your profile. Channel recommendations adjust to your budget. Kill criteria fire faster or slower depending on your stated risk tolerance.

Path	Time	What it asks
Full	~5 min, 11 questions	Technical level, domain, past projects, interests, what people ask you for, communities, content you consume, strengths, audience, constraints, target buyer
Short	~2 min, 5 questions	Technical level and domain, audience, inner circle, constraints, target buyer
Browse	~1 min	Pick 2–3 domains from 20 product categories people actually pay for
Skip	15 sec	One mandatory question, generic recommendations

Four of those answers do specific work: hours per week decides whether a slow content channel can ever ramp, monthly budget decides whether paid acquisition exists at all, risk tolerance sets how fast the kill criteria fire, and target buyer selects which rubric lane every downstream skill uses.

Your profile is a local Markdown file. Nothing is uploaded anywhere.

What a real verdict looks like

This is not a mock-up. It is an abridged run from examples/, which ships with the repo. Two complete validations are in there, unedited, with every source URL and every intermediate file.

Decision Memo: Shadow AI Discovery for SMB, Sold Through MSPs

Verdict: PIVOT — 43/100 · Confidence: medium · 15 sources checked

Why this score. The market is real and the product is buildable. Neither is the problem. Microsoft now bundles Purview into M365 Business Premium, so adequate discovery costs a Microsoft-native client nothing extra, and two vendors already sell per-device shadow-AI discovery to MSPs. Competition scored 22, low enough to trigger the killer-dimension penalty. Meanwhile you have no route to this buyer.

Riskiest assumption. "MSPs will buy a standalone SKU from an unknown solo vendor, rather than waiting for the platform they already resell to ship the same thing."

Test it in 14 days for under $50. Contact 25 MSP owners as a practitioner, not a vendor. Ask whether a client raised AI controls recently, then: at $2 per device, would you buy now or wait for Huntress or Cynomi to add it? Pass: 8 of 25 say buy now, 2 put a card on file. If 15 say they would wait, it fails regardless.

Devil's advocate. The demand data is the strongest in this project and I am discounting it on channel grounds. Businesses have been built into worse markets with better hustle.

That idea had scored 53/100 at candidate stage on trend data alone. Real competitor research moved it to 43, and moved the competition dimension from 55 to 22.

Read both full runs → Two ideas, two verdicts, ten evidence files each, and a command that proves the numbers were not touched by hand:

python tests/validate_memory.py --memory examples/memory
SUMMARY checked=20 errors=0 warnings=0

This keeps happening. Across five full validations in this repository, the competition score fell every time research replaced assumption:

Idea	Estimated	Researched	Drop
Automation security scanner	88	38	−50
Delivery provenance attestation	85	45	−40
Agent governance layer	85	65	−20
GEO category panel	62	32	−30
Shadow AI for MSP	55	22	−33

So the system now caps unresearched competition scores at 45 and says so out loud. That recalibration is in docs/HANDOFFS.md, with the evidence.

Four workflows

Find me an idea · ~10–15 min

Interview → market research → 7–10 scored candidates matched to your actual edge. Ranked by score with a rank label, never a fake verdict, because a candidate scored on partial data has not earned one.

Validate this idea · ~15 min

Ten steps: trend research → competitors → desire → pricing → market size → retention → CAC → score → memo. Out comes a decision memo: verdict, three strengths and three risks with data behind each, the riskiest assumption with a costed experiment, a pre-mortem, a devil's advocate section arguing against the verdict, kill criteria, and sources.

In a hurry? Say "gut check" for the fast path: four dimensions, three minutes, no memo. It can rule an idea out but is arithmetically incapable of telling you to build — which is the point.

Tell me about this market · ~10–15 min

Trend velocity → competitive landscape → TAM/SAM/SOM → how people actually acquire users here. Say "quick look" for trend analysis only.

Should I pivot? · ~10–15 min

Root-cause the weak dimensions, generate 2–3 evidence-backed pivots with projected scores and effort estimates, re-score the best one. Already know what you want to change? Say "what if I charged per client instead of per seat" and it runs a micro-pivot on that single variable.

It knows the difference between consumer and business

Most idea tools assume you are building a consumer app. Every idea here declares who pays, and eight skills switch rubric accordingly.

	Consumer lane	Business lane
Competitors	App Store search, 1★/3★ review mining	G2 and Capterra, pricing pages, Hacker News, operator forums
Pricing	Van Westendorp, category WTP bands, freemium conversion	Contract-value tiers, price anchors, trial-to-paid, rebilling
Distribution	ASO scoring, creator fit, k-factor	Discovery-channel scoring, advocacy fit, channel ceilings
Market size	Community proxy, capture-rate benchmarks	ICP count × contract value × win rate
Retention	D1 / D7 / D30 by category	Monthly logo churn by segment
CAC	Cost per install	Three-case LTV, gross margin, ceiling per channel

Set it once at onboarding, override it per idea.

The methodology, and where it comes from

Technique	What it does	Origin
Multiplicative floor	Any dimension under 25 multiplies the final score down	Models how one fatal flaw actually kills a company
Riskiest Assumption Test	Designs a ≤2-week, ≤$100 behavioural experiment for the single assumption most likely to be fatal	Lean startup practice
Pre-mortem	Assumes the idea failed in 12 months, works backwards to the three likeliest causes	Klein, 2007
Devil's advocate	Three strongest arguments against the verdict, each tied to a data point	Added because memos were too agreeable
Van Westendorp	Four price thresholds to bracket willingness to pay	Price sensitivity meter, 1976
Viral coefficient (k)	k = invites × conversion, classified across named loop types	Standard growth accounting
Triangulated sizing	Three independent estimates cross-checked; disagreement lowers confidence	Bottom-up market sizing

Six dimensions, weighted: demand 20%, competition 10%, monetization 20%, distribution 20%, retention 15%, founder-market fit 15%. Verdicts at 75 (pursue), 55 (test), 35 (pivot), below that drop.

Every benchmark either cites a source or is explicitly labelled a heuristic. There is no third category.

The part that keeps it honest: it tests itself

Prompt systems drift silently. Outputs slowly stop matching the specification, and nothing complains.

python tests/validate_memory.py --check-specs      # specs, workflows, adapters
python tests/validate_memory.py --idea <slug>      # one analysis, end to end
python tests/make_fixtures.py --out /tmp/fx        # 10 fixtures, pass and fail paths

Standard library only. No dependencies. It checks that:

every score's arithmetic recomputes exactly, including the floor penalty and the missing-data discount
every output file matches its contract, with the right fields for its lane
every verdict matches the threshold table
every memo has its required sections, sits in its word budget, and cites at least three URLs
every research file has sources and a real staleness date
no workflow reads a file that no skill produces

Six fixtures must pass; four are deliberately broken and must fail with an exact error count. When a threshold was tightened, a stale fixture caught it within seconds.

What gets saved

Everything is a local file. Nothing is uploaded.

memory/
├── user_profile.md                          your background, reused across sessions
├── market_insights/
│   └── <niche>-<platform>-<YYYY>-<MM>.md    research with a Sources section
└── ideas/
    └── <your-idea>/
        ├── idea.md              competitors.json     pricing.json
        ├── market_size.json     distribution.json    retention.json
        ├── cac.json             desire_scores.json   scores.json
        └── decision_memo.md     ← the thing you actually read

Ideas are never deleted, only marked dropped or paused. A pivot that changes enough to be a new idea gets its own directory with a pointer back to the original, so you keep the record of what you tried.

Your own memory/ is gitignored, so your ideas stay yours. Two finished analyses ship in examples/ so you can see the shape of the output before running anything.

Honestly, how is this different from just asking ChatGPT?

A general chatbot can do any single step here, often well. What it does not do:

	General chatbot	FindMeSaaS
Consistency across ten steps	Re-invents its criteria each time	One rubric, same thresholds, every run
One fatal flaw	Averaged into a friendly summary	Multiplies the score down
Evidence	Usually unsourced	URL and access date, written to disk
Knowing what it doesn't know	Rarely says	Missing dimensions discount the score
Next session	Gone	Files you can re-read and diff
Being wrong loudly	No mechanism	Test suite fails the run
Who you are	Generic advice	Scored against your hours, budget, risk and skills

Use the chatbot for exploring. Use this when you are about to spend six months.

FAQ

Do I need API keys or a paid plan? No. It runs inside a coding agent you already use. Web research uses that tool's built-in search.

Will it ever tell me my idea is good? Sometimes. The two ideas validated in this repository scored 43 and 28, and one of those was chosen deliberately as a control. A pursue verdict needs 75 out of 100 with all six dimensions researched, which is hard on purpose.

Can I trust the numbers? Trust them as far as their sources. Every figure is either cited with a URL or marked as a heuristic, and each analysis carries its own confidence rating and a list of what it could not find.

Does it work for B2B? Yes, and that is the part most idea tools skip. See the lanes table above.

Does it send my ideas anywhere? No. Memory is plain files in the repo. The only network traffic is the web research your agent performs.

What if it is wrong? Open an issue with the idea slug and the dimension you disagree with. Five rubric recalibrations in docs/HANDOFFS.md came from exactly that.

Contributing

See CONTRIBUTING.md. The short version: edit the canonical skill in skills/, leave the three platform adapters alone, and run the harness before opening a pull request.

The most valuable issue you can file is a verdict you think is wrong.

Credits

Built and maintained by Latif Abderrahmane — the business-model rubric lanes, the validation harness and fixture suite, the fast-path, micro-pivot, quick-scan and pivot-lineage workflows, the B2B research prompt, and the scoring recalibrations.