idea-killer

Your AI is a hype man. This makes it a prosecutor.

Two Claude Code skills that put every new business/product/side-project idea on trial - and remember every verdict, so the same bad idea can never cost you twice.

you:    /idea-killer  a local-first budgeting app, one-time purchase, privacy-focused

claude: Standing Prior P1 matches (privacy-premium fallacy, 3 kills).
        Evidence bar raised: this needs someone else's money, not reasoning.

        1. Gap-reality ......... FAIL - no observed demand named, only a held value
        2. Incumbent-patch ..... FAIL - YNAB ships "local vault" in one quarter
        3. Paid-comparable ..... UNPROVEN - none named
        4. Buyer math .......... FAIL - one-time price vs permanent support tail
        ...
        Verdict: KILL - test 4. The monetization flaw is structural: the repair
        is SaaS, and SaaS is thesis-breaking for a no-accounts privacy product.

        Ledger entry LED-014 written. Tag "privacy-premium" now at 4 kills.

Why this instead of a "devil's advocate" prompt

Prompts that argue against your idea are everywhere. They all share the same flaw: they forget. Next week you repackage the same dead idea with a new coat of paint, the prompt evaluates it fresh, and you lose another month.

idea-killer writes every verdict to a persistent, git-auditable ledger:

• IDEA-LEDGER.md - every idea ever evaluated: verdict, the exact test it failed, failure tags, re-entry triggers.
• PITFALLS.md - your standing priors: when three kills share a failure tag, it gets promoted to a named pattern in YOUR decision-making. Future ideas that match a prior face a raised evidence bar - they need external proof (someone else's money or usage data), not reasoning, to survive.

The prompt is the write path. The ledger is the product. It compounds: the more ideas you kill, the sharper your priors, the cheaper every future evaluation.

What makes the evaluation actually adversarial

Eight hard tests, one hard fail = KILL. Highlights:

• Incumbent-patch test - if an existing player ships it as a feature in a quarter, you are not building a product, you are writing their roadmap.
• Paid-comparable test - name a WORSE version people already pay for. No comparable = no market until proven otherwise.
• Distribution test - name the channel and why YOU win it. "Post it and they will come" is a fail.
• Meta-work test - is this tooling about your own process dressed up as a product? (If you build with AI a lot, you know this one.)
• Anti-flattery gate - if the verdict would make you feel good about proceeding, the model must re-score before output.

And the part most validators skip: Repairs. Even on a KILL, the skill names the smallest change that would fix each failed test, and classifies it COMPATIBLE or THESIS-BREAKING. Knowing an idea died of a structural flaw (not a tuning problem) is what stops the zombie rebuild six months later.

Verdicts are KILL / PARK (with a written re-entry trigger) / GATE (objective, pre-registered numbers) / GO (which still requires a < 1 week paid-validation step before any build).

Install

As a Claude Code plugin:

/plugin marketplace add DepthWorks/idea-killer
/plugin install idea-killer@depthworks

Or manually - copy the two skill folders into your global skills directory:

git clone https://github.com/DepthWorks/idea-killer
cp -r idea-killer/skills/idea-killer idea-killer/skills/ledger ~/.claude/skills/

The ledger lives at ~/.claude/ledger/ by default (override with the IDEA_LEDGER_DIR environment variable). On first run the skill offers to initialize it from ledger-template/. Keep the ledger in a git repo: an uncommitted ledger is not auditable, and the skill will remind you.

Usage

/idea-killer <your idea>        put an idea on trial
/ledger have I tried X before   shape-match against past kills
/ledger what are my pitfalls    your standing priors, verbatim

The ledger skill matches on an idea's SHAPE (same buyer, same wedge, same distribution problem), not its wording - a new coat of paint on a killed shape is the same idea.

Design notes

• No scripts, no hooks, no network. Two plain-markdown skills you can read in five minutes before trusting them. Audit them like you should audit every skill you install.
• Token-lean. Written for real-world budgets (works fine on the Pro plan): the judgment lives in the skill file so the model does not re-derive it, and nothing loads until you invoke it.
• Model-agnostic discipline. The execution notes are binding on any model: score PASS/FAIL/UNPROVEN with one sentence of proof each, UNPROVEN counts as FAIL for a GO, and "none named" is the required answer when no comparable or channel exists - never a plausible-sounding placeholder.

License

MIT. The ledger it writes is yours.