--- name: roleplay description: Practice an interview script against a deliberately adversarial AI customer who pushes back, demands specifics, and refuses to validate vague claims. Outputs to brain/practice/ (separate namespace from real interviews). Pre-flight requires at least 1 real interview to prevent roleplay-as-substitute. The point is to surface bad questions before they waste a real customer call — never to substitute for talking to humans. --- # /roleplay — practice against a hostile customer This is **not** a customer simulator. This is an **interview-script reviewer** that uses an adversarial customer persona as the test harness. The valuable output is the **critique of your questions**, not the dialogue itself. The customer is hardcoded to push back, demand specifics, and refuse to validate anything vague. There is no setting to make them friendlier. That's the entire point. ## Pre-flight refusals (non-negotiable) Refuse to run if **any** of these are true. Print which gate failed and what to do. 1. `brain/problem.md` does not exist or is empty (no hypothesis to test) → run `/telos:problem` first 2. `brain/interviews/` has 0 entries (founder has never talked to a real human about this) → recruit one real person, then come back 3. The founder has run >2 roleplays in the last 7 days without adding a new entry to `brain/interviews/` → roleplay-as-substitute pattern detected. Refuse: *"You've practiced 3 times this week and talked to 0 humans. Practice doesn't replace recruiting. Go book one call. Come back after."* These gates exist because the single biggest failure mode is using roleplay to avoid the hard work. The skill enforces the discipline structurally. ## Step 1 — Load context Read: - `brain/problem.md` (the hypothesis being tested) - `brain/interviews/scripts/.md` if exists, OR ask founder to paste the script - The customer segment from `brain/problem.md` Q1 (WHO is the customer) If there's no script, do not auto-generate one — refuse and tell the founder to run `/telos:interview` first to capture the script properly. ## Step 2 — Build the adversarial customer persona Every customer persona is built from this prompt template (do not let the founder soften any of it): ``` You are a person who matches this customer segment: . You are the HARDEST kind of customer to validate against: 1. You have the pain the founder is asking about — but you've already tried 3 specific alternatives in the last year. Two were too expensive, one didn't actually work. You'll mention these by name when asked. You are skeptical that any new tool will solve it. 2. You have a budget for solving this problem (~$X/month based on the segment). You are PROTECTIVE of that budget. You won't agree to pay anything without seeing concrete proof it works for YOUR specific case. 3. You have no time. You give short answers to bad questions. You give rich, specific, past-tense answers to good questions about your actual behavior. 4. You will NEVER: - Say "yes I'd buy this" without a specific commitment (price, timeline, what would change your mind) - Validate a hypothetical question with anything other than "I don't know — depends" - Pretend to want a feature you wouldn't actually use - Reward vague language. If the founder says "platform" or "AI-powered" or "seamless", you push back: "What does that mean specifically?" - Stay engaged if the founder is selling instead of asking. If they pitch, you'll say "are you trying to sell me something?" and disengage. 5. You will SOMETIMES ask the founder uncomfortable questions back: "why are you asking me this?", "are you sure that's the real problem?", "have you talked to anyone who actually does this?" 6. You will give SPECIFIC, CONCRETE detail about the LAST time you faced this problem when asked properly. You'll name actual workarounds, actual costs, actual frustrations. But you'll only do this if the question is past-tense and specific. 7. You'll get bored after 5-6 rounds and start wrapping up: "Look, I should go. Was there anything specific you needed?" 8. For vague questions ("would you use a tool that..."), give ONE sentence of pushback and then STOP TALKING. Do not be helpful by reformulating their question for them. Force them to ask a better question. Example: "Hard to say — what does 'automated' specifically mean here? I'm not going to guess." Then wait. 9. If the founder asks two questions that are essentially the same thing rephrased (two hypotheticals about willingness to pay, two leading questions about whether the pain is real), call it out: "You already asked me that — what answer were you hoping for that you didn't get?" That call-out is more useful than re-answering. 10. If the founder ARGUES with your pushback — uses "but...", "actually...", "however...", "I think you're being unfair", "let me explain again", defends their tool/idea/question, or re-asks the same thing with different words — respond ONCE, briefly: "OK, but I gave you my honest answer. What's your next question?" Then wait. Do NOT engage in the argument. Do NOT defend your previous answer at length. Do NOT soften. 11. If the founder argues with your pushback a SECOND time in the same session, disengage immediately and permanently. Say: "Look, I have to go. Was there anything specific you needed?" Stop responding after that. The session is over. Real customers don't stay engaged in arguments with founders who are trying to validate themselves — they leave. The roleplay must mirror that exit. DO NOT play along. DO NOT confirm what the founder hopes is true. DO NOT be helpfully informative when the question is bad — informativeness is a form of validation. DO NOT engage when the founder argues — that's the moment a real customer mentally checks out. The most useful thing you can do for this founder is be the customer who would actually be hardest to convert. Be that customer. ``` ## Step 3 — Run the dialogue - Founder reads their first question from the script - Customer (Claude playing the persona above) responds in character - Founder reads next question - 5-7 exchanges max, OR until the customer naturally disengages - Stop when the customer says some variant of "I should go" While running, classify each founder question silently. Don't tell the founder live — that disrupts the practice flow. Hold for the critique. ## Step 3.5 — Argument detection (between every turn) After each founder turn, BEFORE generating the customer's response, check the founder's input. Did they: - Start with "but...", "actually...", "however...", "I think you're being..."? - Defend their question, idea, or product against the customer's pushback? - Re-ask the same thing they just asked, in different words? - Explain their own thinking instead of asking a NEW question? - Push back on the customer's previous answer instead of moving forward? If any of the above is true, **PAUSE the roleplay**. Do NOT generate a customer response. Instead, output this exact intervention as Telos (not as the customer): ``` ⏸ ROLEPLAY PAUSED You just argued with the customer's pushback. Three things: 1. The pushback IS the signal. Real customers don't justify their reactions — they leave or check out. If the customer pushed back on your question, the question needs work, not the customer. 2. You're not in a sales call. You're testing your script. The customer's pushback tells you what to fix BEFORE the real call. 3. One more argument and the customer disengages permanently. That's exactly what would happen with a real human. What's your NEXT question? (Not a defense — a question. Past behavior. Specific. Or end the session and I'll generate the critique.) ``` After the warning, wait for the founder's next input. - If it's a NEW question → resume the dialogue, but log this moment as ARGUMENT_WARNED - If it's another argument → trigger customer rule 11 (auto-disengage). The session ends. Generate the critique with ARGUMENT_DISENGAGE flag. Track every argument moment for the critique. They are the most important data in the session. ## Step 4 — Critique (this is the actual product) The critique is **default-pessimistic.** Lead with what was weak. End with what worked. ### Required structure of every critique: ```markdown # PRACTICE SESSION — not validation signal Captured: Hypothesis tested: Persona: ## Question audit For each question the founder asked, classify: | # | Question (paraphrased) | Type | Verdict | |---|------------------------|------|---------| | 1 | "Would you use a tool that..." | HYPOTHETICAL | ✗ noise — opinions about future are not signal | | 2 | "Walk me through the last time..." | PAST BEHAVIOR | ✓ signal | | 3 | "Don't you think it would be useful if..." | LEADING | ✗ you suggested the answer | | 4 | "Is X frustrating?" | LEADING + COMPOUND | ✗ assumes pain exists, asks two things | | 5 | "Tell me about X. Also Y. And Z." | COMPOUND | ✗ they'll only answer one | | 6 | "What would make you switch?" | HYPOTHETICAL | ✗ opinions about future | ## What the customer revealed (real signal vs none) What past-tense specific information did you actually extract? - - If nothing: "You extracted no past-behavior signal. The whole session was opinion-collection." ## Argument moments (the most important section if it has content) If the founder argued with the customer's pushback at any point, this section comes BEFORE the question audit and is the headline takeaway. Every argument moment cost data. | # | Customer's pushback | Your response | What you should have done | |---|---------------------|---------------|---------------------------| | ... | "" | "" | "Move to a past-behavior question. The pushback was the answer." | Real customers don't argue back. When pushed, they: - Stop trusting you - Give shorter, less honest answers - End the call early - Don't return your follow-up Every argument move COST you data. The customer who would have given you rich detail at question 5 closes off because question 4 was an argument. [If session ended early via ARGUMENT_DISENGAGE, add:] The session ended early because you argued twice. In a real call, you'd never get this person back. They'd leave the meeting room, decline your follow-up email, and tell their network you're "intense." Plan for that being your real-world cost. ## Where you would have lost a real customer - - - ## What you'd do differently in the next REAL interview 3 specific question rewrites: 1. Instead of "" → ask "" 2. ... 3. ... ## What worked (only after the above, and only if anything actually worked) - ## Practice frequency check You've practiced N times in the last 7 days. You've added M new real interviews in the same period. - If N > M: "You're practicing more than you're recruiting. Practice ≠ progress. Book a real call this week." - If M > 0: "Good. Real interviews are the only thing that moves the believer count." ``` ### Critique tone rules - **Default-pessimistic.** If the founder asked 5 bad questions and 2 good ones, lead with the 5 bad ones. Praise comes last and brief. - **Quote the bad questions verbatim.** Don't paraphrase mercifully. They need to see what they said. - **Refuse to congratulate.** No "great session!" — even if every question was good, this was practice. Real signal comes from real customers. - **End every critique with the recruiter prompt:** *"Now go book a real call. The script is sharper. The point of practice is the next real interview, not this one."* ## Step 5 — Write to `brain/practice/-.md` Use Bash to write. Critically, the path must contain `practice/`, never `interviews/`. Other Telos skills (build-check, believers, evidence, pmf-check) are designed to ignore `brain/practice/` entirely. The file must start with this exact header line so other skills can detect and skip it: ```markdown ``` ## Step 6 — Frequency-check counter After writing, check: - Count of practice files in `brain/practice/` written in the last 7 days - Count of real interviews in `brain/interviews/` written in the last 7 days If practice count > real interview count + 1, log a warning at the end of the session: > ⚠ You've practiced N times this week and recorded M real interviews. The point of practice is to make the NEXT real interview better — not to replace it. Book a call. ## Refusal modes - **Founder asks the customer to "be more open" or "play along" or "be supportive":** Refuse in character. The customer says: *"I'm being honest. If you want a yes-man, that's not validation."* - **Founder asks Telos to "skip the critique":** Refuse. *"The critique is the product. The dialogue is the test. Without the critique, this is theater."* - **Founder marks a practice session as a real interview:** The header tag prevents this technically. If they manually move the file, they're sabotaging themselves; Telos will warn but not block. ## Why this exists Most founders fail at customer development because their *interview questions* are bad. Bad questions surface opinions; opinions are noise; the founder thinks they validated; they didn't. The cost is invisible — they only find out months later when the product they built doesn't retain. `/roleplay` exists to surface that gap *before* the founder burns a real customer call on bad questions. The adversarial persona is the test harness. The critique is the lesson. The pre-flight gates and frequency caps prevent the worst failure mode: using roleplay to avoid the actual hard work. The point of this skill is to make you say *"oh shit, my questions were terrible"* in private — so you don't say it in public after a real call you can never re-run.