Make Claude say “I’m out” before you waste six months building.

You have an app idea. You ask an AI assistant whether it's good. It politely tells you the opportunity is promising, gives you a feature roadmap, and offers to help you start coding. That feels great. It is also a terrible substitute for finding out whether anyone will actually pay.

The interesting question isn't “Can I build this?” anymore. For many software ideas, the harder questions are who needs it, why now, why they would pay, how you would reach them, and what happens when your initial assumptions are wrong.

That's the appeal of shark-skill, a free Claude Code skill made by Alex Chen (@nocodealex). It lets you pitch a business idea to five simulated “sharks” instructed to look for reasons to say no. Instead of returning another encouraging paragraph, the workflow tries to expose the unanswered questions you should investigate before committing months to a build.

What is the SHARK skill?

/shark is a Claude Code skill, not a special investor model and not a real investment platform. You give it your idea and any evidence you have. It orchestrates a structured mock pitch: multiple investor personas ask uncomfortable questions, a separate founder persona responds using your supplied facts, and the investors decide whether they would make an imaginary offer or walk away.

The output is a Markdown board containing each investor's decision, their reasoning, what would change their mind, and the questions your pitch couldn't answer. The project describes that report as BOARD.md, generated inside a local .shark/ folder.

What actually happens when you run it?

Meet the sharks

These are the five perspectives described for the default panel. The project also has a larger pool of ten investor personas, and the selection can be adjusted with flags.

These are different prompts and roles, not genuinely independent human judgment. Their value is in forcing you to consider a wider range of failure modes than you might think of alone.

Before installing: what you need

• Claude Code installed and usable on your machine. This is a Claude Code command, not something you run in an ordinary browser chat by typing /shark.
• Access to Claude usage. The skill's code is free and MIT-licensed, but its multi-agent runs consume usage from your Claude plan or setup.
• A project folder where the skill can write its .shark/ results.
• A reasonably specific pitch. The more concrete your customer, pricing, costs and evidence, the harder it can interrogate your business.
• Permission to install third-party code. Inspect the repository and its scripts before asking an agent to install and execute them.

The repository says the default five-investor run uses 11 sub-agent calls: five sets of questions, one founder pass, and five decisions. That's why the run can be more expensive in usage than one ordinary Claude prompt. The maintainer also suggests using accept-edits mode (Shift+Tab) if you don't want repeated confirmations for the files it writes; review what the skill does before enabling less restrictive permissions.

Install the free skill

Option A — Ask Claude Code to install it

Open Claude Code in your working directory and paste the following. This is the installation method provided by the repository. Inspect what Claude proposes to do before allowing file changes or execution.

https://github.com/alexyc9381/shark-skill
Install this skill, then confirm /shark works.

Start a new Claude Code session if the command isn't discovered immediately, then verify that /shark is available. Installing from a public repository does not guarantee the code is safe; you remain responsible for reviewing it.

Option B — Manual install

If you prefer to inspect the contents yourself, clone the project and copy its skill directory into your local Claude skills folder.

git clone https://github.com/alexyc9381/shark-skill
mkdir -p ~/.claude/skills
cp -r shark-skill/skills/shark ~/.claude/skills/

Option C — Skills CLI

npx skills add alexyc9381/shark-skill

Option D — Claude Code plugin marketplace

/plugin marketplace add alexyc9381/shark-skill
/plugin install shark-skill@shark-skill

Pick one installation approach. You don't need to use all four. If the author changes the repository layout or Claude Code changes its plugin behavior, follow the current repository installation section rather than relying on an old screenshot.

Run your first pitch

Start with an idea you can describe in plain language. Here's a fictional example written for this article—not a claimed result from a real run.

/shark A lightweight API monitoring tool for solo developers.
It alerts users when endpoints fail and keeps simple uptime history.
Target customer: indie developers running 1–5 projects.
Proposed price: $9/month.
Traction: 14 free users, 0 paying users.
Current costs: roughly $18/month.
Main question: what would make this worth paying for?

Notice that this prompt says who the customer is, how many users exist, what they pay (or don't), what it costs to run, and the biggest unresolved question. That is far more valuable than asking the system whether “an AI app for developers” is worth building.

Make the panel harder or reproducible

The skill's public example uses a local hot-sauce subscription box and a two-shark session. According to the repository, neither investor offered a deal and the board produced six homework items. You can inspect the example directory rather than taking a fabricated demonstration at face value.

Three copy-paste pitch templates

These are my own reusable templates, inspired by the skill's interface. Replace every bracket with real information; explicitly write “none yet” when you don't have proof. Unknowns are useful inputs to this exercise.

For a SaaS or developer tool

/shark [My product] helps [specific customer] solve [painful problem].
Today they solve it by [current workaround or competitor].
I plan to charge [price] through [pricing model].
Evidence: [interviews, signups, revenue, usage, or 'none yet'].
Acquisition channel: [channel and reason I expect it to work].
Costs: [fixed and per-customer costs].
Biggest unknown: [what I most want challenged].

For a creator business

/shark I create [type of content] for [specific audience].
Audience: [followers and typical reach].
I want to sell [product] at [price].
Evidence of demand: [buyers, survey replies, email list, or none].
How I would reach customers: [distribution].
Challenge the willingness to pay and fulfillment costs.

For a service, consultant or agency

/shark I offer [specific service] to [specific customer].
Price: [fixed fee or retainer].
Delivery cost and time: [hours, tools, contractors].
Current proof: [clients, outcomes, testimonials, or none].
Competition: [who else offers it].
Tell me what a skeptical buyer would reject first.

How to read the resulting BOARD.md

Resist the temptation to celebrate when three sharks say IN or give up when all five say OUT. Both outcomes come from a role-play. Treat the report as an ordered collection of objections, not a valuation or a probability of success.

A fictional example of useful criticism

Suppose your API monitoring tool has fourteen free users and zero paying customers. The panel might ask why a developer would pay $9 when free monitoring tools already exist. That isn't proof the product will fail. It's a testable uncertainty: what specific inconvenience would convince this group to pay?

You could interview current users, offer a paid plan to a small number of them, or manually provide an alerting service to see whether the workflow is valuable before investing in a larger codebase. The win is that the question becomes concrete.

Turn investor objections into a real-world test

Here's the step that matters more than the role-play itself. Use the board to decide what to learn next—outside Claude.

After the panel, use this separate follow-up prompt:

Turn the unanswered questions in the board into a 7-day validation plan.
For each risk, give me:
- one falsifiable assumption
- the cheapest real-world experiment to test it
- the success/failure threshold I should define BEFORE testing
- the evidence I need to collect
- what decision to make if the test fails
Do not invent customer evidence or revenue.

Example: for a $9/month API monitoring tool, you might decide in advance to contact 20 developers who actively operate side projects, secure five genuine product interviews, and ask for a paid pilot from those who report a specific monitoring problem. Those numbers are illustrative experiment settings, not universal success thresholds. Choose thresholds that fit your business and sample.

Where this is useful—and where it breaks

What it is good for

• Blind-spot discovery: bringing up questions you've been avoiding because you enjoy building.
• Pitch preparation: practicing answers around demand, margins, differentiation and acquisition.
• Structured documentation: walking away with a prioritized record rather than a pleasant chat.
• Pre-build triage: identifying cheap experiments before sinking weeks into a feature set.
• Team discussions: giving collaborators a shared list of assumptions to debate.

What it cannot do

• Validate demand: a model cannot replace customers using or paying for something.
• Predict fundraising: fictional offers from Claude are not evidence that investors will fund you.
• Guarantee independence: five role prompts can still share the same underlying model biases.
• Verify your numbers: it relies on the information provided; inaccurate inputs produce unreliable critiques.
• Guarantee privacy: the project keeps report files in a local folder, but prompts still go through your configured Claude service. Don't paste confidential business details you cannot share.

A practical one-weekend workflow

If you're the kind of person who can build an MVP in two days, here's a more disciplined way to use those two days.

Frequently asked questions

The takeaway

Most founders don't need another AI assistant telling them they're building the next billion-dollar company. They need someone—or something—to ask why a specific customer would care, what evidence exists, and what would make the whole thing fail.

/shark is a useful rehearsal, not a verdict. Its best output isn't a simulated investment offer. It's the list of questions you're now responsible for answering in the real world.