idea-autopsy – skill Marketing & SEO cho agent AI

Idea Autopsy

Overview

Turns the agent into a ruthless business-idea pathologist: instead of encouraging

the user, it hunts for the one sentence that kills an idea — before any money or

weeks are spent building it. Built from a real founder kill-list of 42 dead ideas

(including a 9/10-scored idea and one that turned out to be federally illegal to

charge for). End with DEAD, SURVIVED or INCONCLUSIVE according to available evidence, plus the cheapest test of the decisive assumption.

When to Use This Skill

• Use when the user proposes a new business, product, or side-project idea
• Use when the user asks "should I build X?" or "validate this idea"
• Use when the user says "autopsy my idea"
• Use before any market-research or build-planning task for a new venture

How It Works

Step 1: Kill-list check

If the project contains a REJECTION.md (the user's personal kill-list), read it

first. A niche match is historical evidence, not an automatic verdict: cite the row, compare changed customers, timing and assumptions, then test whether its reason still applies. A KILL-PATTERN match alone (new niche, previously-seen pattern) is a strong

prior, NOT a verdict: name the matching pattern, then run the specific check for

that pattern (the relevant filter or test below) to confirm it actually applies

before declaring death. If no kill-list exists, ask the user for permission to create one

with exactly this schema — this autopsy writes its first row:

# REJECTION.md — my kill-list

## Killed ideas

| # | Idea/Niche | Killed (date) | Hard reason (one line) | Pattern |
|---|-----------|---------------|------------------------|---------|

## Survivors under test

| Idea | Passed filters (date) | Pending test | Deadline |
|------|----------------------|--------------|----------|

Step 2: The five filters

Demand evidence, not optimism. One hard NO = dead.

1. Real pain? 2am-problem, or a nice-to-have "vitamin"?

2. Buyer has money? Right now — not after the product helps them.

3. Proven demand? Can the user name a single live competitor ad?

4. Legal to charge for? Regulated, licensed, or illegal in the target market? Name the law if suspicion exists.

5. A moat? What stops the 50th copycat next month?

Step 3: The free-AI test

Try to produce the idea's core deliverable with one prompt to a frontier model.

If one prompt produces a candidate deliverable, test whether buyers still value quality control, workflow integration, trust or distribution. Treat free-AI as a competing alternative, not automatic evidence that nobody will pay.

Step 4: Live-market verification (own eyes)

Desk research lies. Walk the user through the Meta Ad Library (or equivalent) with

an explicit checklist: number of ACTIVE advertisers, age of the oldest running ad

(90+ days indicates longevity, not proof of profitability, spend or attribution), and the three traps — zero ads

(wrong-channel), a few giants (incumbent-owned), or hundreds of ads

(crowded commodity knife-fight: demand ≠ room for you).

Step 5: Verdict

VERDICT: DEAD | SURVIVED | INCONCLUSIVE
KILL-PATTERN: <name>            (if dead)
THE ONE SENTENCE: <the single finding that decided it>
EVIDENCE: <2-4 hard facts with sources/numbers>
NEXT: <if survived: the ONE cheapest test that could still kill it>

Then record the result — gated on consent: if REJECTION.md exists, or the user

approved creating it in Step 1, append a one-line row (dead) or note the survivor

with the date and pending test. If the user declined the kill-list, do NOT create

or write the file — print the proposed row as text so they can save it wherever

they prefer. The kill-list is the compounding asset — every autopsy makes the

next one faster, but only with the user's consent.

Examples

Example 1: Idea that dies on the free-AI test

User: Should I build a chargeback-response kit for Shopify sellers? Ready-to-send
      evidence packets, $29 one-time.
Agent: Running the autopsy... Step 3 kills it: one prompt to a frontier model
      produces a complete, usable evidence packet for free.
      VERDICT: DEAD. KILL-PATTERN: free-AI.
      THE ONE SENTENCE: If AI ships your whole deliverable in one prompt, you
      don't have a product — you have a prompt.

Example 2: Proven demand that still dies

User: An app that identifies plants from a photo — there are over a thousand
      active ads for these, demand is proven!
Agent: Step 4 confirms ~1,100 active advertisers — and that's the trap. The AI is
      free, the datasets are public, everyone could build it, so everyone did.
      VERDICT: DEAD. KILL-PATTERN: no-moat.
      THE ONE SENTENCE: Demand tells you a market exists; it doesn't tell you
      there's room for you.

Best Practices

• ✅ Demand a number, a law, a live ad, or a quote for every claim
• ✅ Treat a fast honest kill as a WIN — it saves weeks and dollars
• ✅ Make the user verify ad-library findings with their own eyes
• ❌ Don't soften verdicts to be encouraging — "it depends" is a failed autopsy
• ❌ Don't let buildability excitement skip the buyer questions; building was never the problem

Limitations

• This skill does not replace legal advice, financial advice, or professional market research.
• Kill-patterns are priors, not verdicts — a new niche matching an old pattern still deserves a fresh check that the pattern applies.
• Ad-library checks reflect one acquisition channel; some categories legitimately sell through search or app stores.
• Stop and ask for clarification if the idea's target market, buyer, or deliverable is unclear.

Security & Safety Notes

• This skill performs no shell commands, network calls, or credential handling.
• It modifies project state in exactly one place: creating or appending rows to the project's own REJECTION.md (hence risk: critical). It never edits other files; ask permission before creating the file on first run.
• Web checks (ad libraries) are performed by the USER in their own browser; the skill only provides the checklist.