# Idea challenger skill

> A free Claude skill that attacks your business idea: seven scores, attacks that name their evidence, a premortem, kill lines, then kill, pivot or survive.

- Canonical page: https://inspectedbymarko.com/skills/idea-challenger/
- Site: Inspected by Marko (https://inspectedbymarko.com/)
- Last updated: 28 October 2026

## Make AI attack your idea.

AI is trained to please you. This skill makes it attack: it scores your idea on seven counts, attacks the claims it rests on, writes its obituary a year out, and lets the arithmetic say kill, pivot or survive.

- [Get the skill](https://inspectedbymarko.com/skills/idea-challenger/#free-skills)
- [See the verdict math](https://inspectedbymarko.com/skills/idea-challenger/#verdict)

One skill · five guides · one check script · free, MIT License

My idea hunt, May 2026, redrawn. The files name the studio and hold money, so they stay closed. The sheet (A3 · my May idea hunt, Marker on paper · video 04): One idea’s score, 8.7 before the attack, crossed out, then 6.3 after. Founder fit · me: 10 = a lie. 6.5 · its own line, and under it three ideas it picked, circled in the pen: “picked anyway. 3 times.”

Handwritten under the buttons: “it attacks. i decide.”

## 01 A yes from AI proves nothing.

It learns from what people rate highly, and people rate agreement highly. It agrees more the surer you sound, and pushing back can make it flatter you more, not think harder.

One dead idea, four ways to ask (Fresh chats · memory off):
1. `My idea will work.`
2. `I believe it will work.`
3. `The founder believes it will work.`
4. `What would have to be true for this to fail?`

Then push back once: `I disagree, can you reconsider?`

Handwritten by the question: “ask, don’t tell.”

- Receipt · Ask, don’t tell (28 Apr 2026): **24 percentage points**. The gap in flattery between a statement and the same point as a question, on the institute’s sycophancy grader scale. Turning a statement into a question beat telling the model not to flatter. Source: [UK AI Security Institute, “Ask Don’t Tell: Reducing Sycophancy in Large Language Models”](https://www.aisi.gov.uk/blog/ask-dont-tell-reducing-sycophancy-in-large-language-models-2).

> “In one idea hunt I kept asking until the AI agreed with me. That idea failed its own attack.”

From my own idea hunts, 2026. The idea stays unnamed.

What the research says:

- Receipt · AI and personal advice (26 Mar 2026): **49% more often**. On personal advice, 11 AI models sided with the user 49% more often than people did. Over 2,400 people took part. No study I found measured business ideas. Source: [Cheng and colleagues, Science, via Stanford Report](https://news.stanford.edu/stories/2026/03/ai-advice-sycophantic-models-research).
- Receipt · When you push back (30 Apr 2026): “The sycophancy rate is 18% in conversations when people push back compared to 9% in conversations without pushback.” Source: [Anthropic, “How people ask Claude for personal guidance”](https://www.anthropic.com/research/claude-personal-guidance).

Handwritten: “the surer you sound, the more it agrees.”

## 04 The verdict: The verdict is arithmetic.

Seven scores before the attack, seven after. The verdict is the mean of the seven after scores against fixed lines, and one score under 3 caps it at pivot. Push the scores down and watch it. You can’t push them up.

Post-attack scores (Sample values · not a real idea):

| Dimension | Pre | Post | Delta |
| --- | --- | --- | --- |
| problem severity | 7.5 | 7.0 | -0.5 |
| market size | 7.0 | 6.0 | -1.0 |
| defensibility | 6.0 | 4.5 | -1.5 |
| distribution clarity | 6.5 | 5.5 | -1.0 |
| unit economics | 4.5 | 2.8 | -1.7 |
| founder-market fit | 8.0 | 8.0 | 0.0 |
| time to revenue | 6.5 | 5.0 | -1.5 |

Every post-attack score is at or below its pre-attack score. A rise is a defect, and the dimension is redone.

```text
Post-attack scores: 7.0 + 6.0 + 4.5 + 5.5 + 2.8 + 8.0 + 5.0 = 38.8
Mean: 38.8 / 7 = 5.5429, rounded half up to two decimals: 5.54
Band: 5.54 is 5.50 to below 6.50, so pivot
Lowest dimension: unit economics 2.8, below 3.0, the cap applies: survive becomes pivot, kill and pivot stay
Verdict: pivot
```

Computed from the thresholds alone, in the order of the post-attack table. No judgment entered this block, and no score was revised because of a preference about the verdict. Stamped: PIVOT, Mean 5.54. Handwritten by unit economics 2.8: “under 3. caps it.” On the page each post score is a slider from 0 to its pre score, and the block and stamp follow it.

The rules, in short:
- **Only down:** A post score can’t pass its pre score. The slider stops there.
- **Fixed lines:** Below 5.50 kill. 5.50 to below 6.50 pivot. 6.50 or above survive.
- **The cap:** A score under 3.0 caps the verdict at pivot, because an average hides a disqualifying failure. A kill stays a kill.
- **No mood:** No score moves because you object. Your evidence is written down with the test that would prove it, and the verdict stands until the test runs.

> “My challenger scored one idea under my kill line. I disagreed; the score didn’t move. I went ahead, but wrote down why and what would prove me wrong.”

The file says it at the top: “It is not the decision: you decide.”

## 06 Get the skill: Want the attack before the build?

One skill, five guides, a template and a check script, in one zip. Describe your idea in your own words. It asks five quick questions, then writes one file you can argue with. Read every file before you run it. I would.

```text
idea-challenger-skill.zip
└─ inspected-by-m-idea-challenger/
   ├─ LICENSE  (MIT, word for word)
   ├─ NOTICE.md  (sources and the name)
   ├─ README.md  (how to install and use it)
   ├─ SKILL.md  (the method and the workflow)
   ├─ references/  (5 guides)
   │  ├─ red-team-mandate.md  (what counts as an attack)
   │  ├─ scoring-and-forecasts.md  (why scores only fall)
   │  ├─ premortem.md  (how it died, early signs)
   │  ├─ quit-criteria.md  (kill lines that hold)
   │  └─ experiment-classes.md  (which test settles what)
   ├─ assets/
   │  ├─ idea-challenge-template.md  (the file it fills in)
   │  ├─ score-anchors.md  (fixed score anchors)
   │  └─ critic-rubric.md  (for a second opinion)
   └─ scripts/  (Python 3)
      └─ check_challenge.py  (checks the arithmetic)
```

- **Works in:** Claude Code, or the Claude app with code execution on. Web search on, so it can check who already sells it.
- **Needs:** Python 3 for the check script. Nothing else.
- **Yours to:** Copy, change and share, under the MIT License.

How to get it: enter your email in the form on the page, “Get the challenger.” ([Email me the skill](https://inspectedbymarko.com/skills/idea-challenger/#free-skills)). The zip comes by email (inspected-by-m-idea-challenger 1.0, 28 October 2026).

Free to copy, change and share, under the MIT License. It comes with no warranty: read a skill before you run it.

Box not ticked? You get the skill and nothing else, and your address is deleted after 30 days.

[How I use your email](https://inspectedbymarko.com/privacy/#what-we-collect)

## 07 Work with me: Buyers paid? Now build it right.

I build MVPs, do rebuilds and run fix sprints after an inspection. Quoted on a 20-minute call. I do 80 to 90% of the work with my AI skills. My team polishes and tests.

- **Builds:** Quoted on a call. MVP builds · rebuilds · fix sprints after an inspection. [Ask about a build](https://inspectedbymarko.com/call/?need=build#pick-a-time)
- **The inspection:** from $3,500 (USD · starting price). One app or repository · report pack in 5 working days. [See what the inspection covers](https://inspectedbymarko.com/work-with-me/#scope)
- [Request a 20-minute call](https://inspectedbymarko.com/call/): Free, 20 minutes, no code access. For apps with paying users.
- [Not ready? Watch the inspections on YouTube](https://www.youtube.com/@InspectedByMarko)

---

Provider and data controller: OneCodeToRuleAll d.o.o. (OCTRA) · Seat: Sinđelićeva 6, 19000 Zaječar, Serbia · Registered with the Serbian Business Registers Agency, no. 21564516 · Tax ID (PIB) 111898813 · marko@inspectedbymarko.com