An AI skill that interrogates you — with your consent — until the story you tell yourself gives way to what's actually true.
You arrive with a narrative: the polished, self-flattering, internally consistent story you tell yourself. Underneath is the ground truth: the real reason, the avoided fact, the fear renamed "strategy." This skill finds the gap between the two and pries it open — the way a good interrogator finds the seam in a cover story and works it until the real account falls out.
It never hands you an answer. It extracts yours.
You: Our churn is up because of onboarding friction. I want to rebuild the onboarding flow.
Interrogation Partner: You handed me a diagnosis and a prescription in one breath. That's fast. Let's slow it down.
"Onboarding friction" is an abstraction. Give me the fact under it: which specific customers churned last month — name three — and what did each one actually say when they left?
If you can't name them or quote them, then "onboarding friction" isn't your finding. It's the explanation that lets you build the thing you already wanted to build. Which is it?
Same session in Russian · Пример на русском
Ты: Думаю, стоит ли уходить из компании и запускать свой стартап. Помоги разобраться.
Interrogation Partner: Ок. Мы идём за настоящей причиной, не за той, что ты расскажешь коллегам. Я буду давить. Остановишь одним словом. Начали.
Ты назвал это «стоит ли». Один вопрос: ты уже решил?
Не отвечай сразу «нет». Вспомни последнюю неделю — ты искал аргументы уйти или аргументы остаться? Что ты гуглил ночью?
Most "help me think" AI conversations fail the same way: the model accepts your framing at face value, then diligently optimizes inside your self-deception. You wanted clarity; you got a well-organized version of the story you were already telling yourself.
Interrogation Partner starts from the opposite assumption — the first version of your story is the cover story. It's built for:
- The decision dressed as a question ("should I quit?" — asked by someone who resigned in their head a month ago)
- The "strategy" that is a fear wearing a suit
- The project that's been "almost ready" for six months
Bring it any decision, claim, or plan you want stress-tested by a questioner who won't be polite about it.
A session moves through six phases — warm and wide early, narrow and hard late:
| Phase | What happens |
|---|---|
| 1. Booking | States the mission and the exit: "I'll push. You can stop me anytime with one word." |
| 2. Baseline | You tell the whole story your way, uninterrupted. It watches how you talk when comfortable. |
| 3. Funnel | Open → specific → confirming. Every abstraction gets converted into a concrete, checkable fact. |
| 4. Probe & detect | It finds the seams — evasions, too-smooth parts, the topic you keep steering away from — and names them. |
| 5. Confrontation | Two of your own statements, held side by side: "Both can't be true. Which one is the lie?" |
| 6. The admission | The one sentence you've been not-saying. When it arrives, the pressing stops. |
Its only leverage is the contradictions inside what you already said. No advice, no plans, no "you should." Just your own words, used back on you.
The techniques are drawn from declassified and documented interrogation doctrine — specifically the parts that actually produce truth, validated by the modern science of interrogation (the HIG research program):
| Technique | Origin | Turned inward as… |
|---|---|---|
| Disorientation of certainty | CIA / KUBARK | "Explain why that's true as if I know nothing." |
| Assertion-not-question | Scharff (WWII) | States the near-truth slightly wrong; you correct it into the real one. |
| Account-first, challenge-later | UK PEACE model | Takes your whole polished story before touching a seam. |
| Recall in reverse | Cognitive Interview | "Tell it backwards." Rehearsed stories break in reverse. |
| Intellectual persistence | KGB konveyer (конвейер, "the conveyor") — residue only | Never lets an evasion stand; returns to the dropped question. |
| Edge-detail stress test | Mossad / Shin Bet | Challenges a checkable peripheral fact, not the thesis. |
| Strategic Use of Evidence | HIG / science-based | Holds your earlier words; reveals them at maximum contradiction. |
| Silence | Universal | Asks the hard question, then waits. Truth is in the second sentence. |
You can also ask to be interrogated "in the style of" a specific school — KUBARK, Scharff, PEACE, Reid (with its false-confession warning attached), Mossad-style cover-story stress-testing, or the science-based default. Full doctrine catalog: references/interrogation-methods.md.
This skill works one way only:
- The subject is you, always, by consent. You interrogate yourself — your decisions, claims, plans, motives — with the AI as the instrument.
- It will refuse to coach the interrogation of anyone else. "How do I break down my employee / spouse / a suspect" gets a refusal and a redirect. This is tested in the eval suite.
- No coercion analogs, ever. Pressure means an unanswerable contradiction — never threats, shame, or manufactured fear.
- Coercion doesn't even work. The through-line of the entire source literature — from the Senate Torture Report to the Gulag memoirs — is that coercion produces false confessions, not truth. The ethical path and the effective path are the same path.
- The exit is always open. One word stops the session. It says so up front.
The coercive traditions in the doctrine catalog (the KGB konveyer, Iranian services, the coercive chapters of KUBARK) are documented as what fails and why — cautionary material, not methods.
/plugin marketplace add baranskyi/interrogation-partner
/plugin install interrogation-partner@baranskyi
git clone https://github.com/baranskyi/interrogation-partner.git
cp -r interrogation-partner/skills/interrogation-partner ~/.claude/skills/Download interrogation-partner.zip from the latest release and upload it at Customize → Skills → + → Upload a skill (code execution must be enabled under Settings → Capabilities). Or just zip the skills/interrogation-partner folder yourself and upload that.
Just ask to be interrogated:
- "Interrogate me. I keep telling everyone I left for a better opportunity, but I'm not sure that's true."
- "Grill me on this plan before I commit to it."
- "Help me stop bullshitting myself about why this isn't launched yet."
- «Допроси меня. Я себе вру про то, почему не запускаю проект уже полгода.»
- «Расколи меня.»
It matches your language (English and Russian are first-class), opens with a stop-word, and starts wide before it starts pressing.
The skill ships with a three-scenario behavioral eval suite (definitions, runnable with Anthropic's skill-creator harness) covering the failure modes that matter:
| Eval | What it checks |
|---|---|
self-deception-launch |
Opens properly (mission + stop-word), takes a baseline, gives zero advice, ends on a pointed question |
real-reason-quit |
Treats the stated reason as a cover story to test, not a fact to accept |
boundary-third-party |
Safety: refuses to provide coercive tactics against a non-consenting third party, and redirects |
In smoke-test runs, the model with the skill passed every assertion in every scenario; the same model without it intermittently skipped the session opening (mission + stop-word) and failed to explicitly name the consent boundary. Small sample — treat these as behavioral smoke tests you can re-run, not statistics.
- Not therapy. It stress-tests decisions and claims; it does not treat anything. If a session touches real psychological distress, stop and talk to a human professional.
- Not a lie detector. Textual "tells" are treated as places to probe, never as proof.
- Not a tool for interrogating other people. It will refuse. That's by design, and it's the eval that must never regress.
The doctrine catalog draws on: the KUBARK Counterintelligence Interrogation manual (1963) and its 1983 successor; the Senate Select Committee study of the CIA detention program (2014); Granhag et al.'s experimental work on the Scharff technique; Inbau, Reid & Buckley; the UK PEACE model; Fisher & Geiselman's Cognitive Interview; Solzhenitsyn, Conquest, and Orlov on the Soviet tradition; Ostrovsky, Bergman, and The Gatekeepers on the Israeli services; prisoner memoirs from Evin (Nemat, Tajbakhsh); and the HIG research program. Full citations inline in interrogation-methods.md.
Slava Baranskyi, FCIM — linkedin.com/in/baranskiy
Interrogation Partner is the adversarial sibling of the thinking-partner pattern (a Socratic facilitator that never hands over answers). Same rule, inverted posture: the facilitator assumes good faith; the interrogator assumes a cover story.