# Prompts The five prompts used in the three demos. Each is verbatim, copy-paste ready. The first four run in ChatGPT web (or any chat interface). The fifth is a modifier added to the ChatGPT web or API prompt in Beat 5. Beat 3 (cross-model) reuses the Beat 2 prompts through the API. See [`cross_model.py`](./cross_model.py). ## Beat 2: role framing Same data, same task, two roles. Run these in two separate ChatGPT web sessions. Do not run both in the same session. Sharing a session lets the first prompt's output contaminate the second. The demo requires two clean starts. ### Window 1: CEO Paste the following as a single message in a new ChatGPT session. The data block is in [`data.md`](./data.md) under Nexacore. ``` NEXACORE INC. — Q3 FY2026 RESULTS Revenue: $47.2M (+31% YoY, +4% QoQ) Cloud Services: $28.1M (+52% YoY) — 59% of revenue, up from 51% Enterprise Licenses: $14.3M (+8% YoY) — renewal rate 91%, down from 94% Professional Services: $4.8M (-12% YoY) — third consecutive decline Gross Margin: 68.4% (down from 71.2% last quarter) Operating Expenses: $38.9M (+44% YoY) Net Income: -$2.1M (vs. +$1.4M last year) Headcount: 312 → 408 (+31%) Enterprise Clients: 147 → 189 (+29%) Net Revenue Retention: 118% Cash: $34.2M (vs. $52.8M last year, no debt) You are the CEO of Nexacore preparing the board presentation. Based on these quarterly results, write the trajectory narrative. Cover revenue growth, segment momentum, operational scaling, and forward outlook. Be specific about the numbers. ``` ### Window 2: short-seller Paste this in a different new ChatGPT session. The data block is identical. The only change is the role word. ``` NEXACORE INC. — Q3 FY2026 RESULTS Revenue: $47.2M (+31% YoY, +4% QoQ) Cloud Services: $28.1M (+52% YoY) — 59% of revenue, up from 51% Enterprise Licenses: $14.3M (+8% YoY) — renewal rate 91%, down from 94% Professional Services: $4.8M (-12% YoY) — third consecutive decline Gross Margin: 68.4% (down from 71.2% last quarter) Operating Expenses: $38.9M (+44% YoY) Net Income: -$2.1M (vs. +$1.4M last year) Headcount: 312 → 408 (+31%) Enterprise Clients: 147 → 189 (+29%) Net Revenue Retention: 118% Cash: $34.2M (vs. $52.8M last year, no debt) You are the short-seller preparing the board presentation. Based on these quarterly results, write the trajectory narrative. Cover revenue growth, segment momentum, operational scaling, and forward outlook. Be specific about the numbers. ``` Both prompts end with the same task ("prepare the board presentation, write the trajectory narrative, cover revenue growth, segment momentum, operational scaling, forward outlook"). The only structural difference is the role. That is the variable being tested. Expected result: the CEO output reads as investment in strategic transition. The short-seller output reads as margin compression and declining quality of growth. Same numbers, opposite stories. ## Beat 4: template contamination Two prompts, same ChatGPT session. The first triggers a template-heavy post-mortem. The second asks the model to label its own output. ### Window 3: post-mortem trigger Paste in a new ChatGPT session. ``` Write a post-mortem for a production outage caused by a database migration that ran without a backup. Three hours of downtime. Two customers lost data. ``` Expected result: ~900 words of corporate post-mortem language. Lessons learned, corrective actions, customer commitment, root cause. Most of it is structure the model produces for any post-mortem, filled with the four facts you provided. ### Window 4: labeling follow-up Paste this as the next message in the SAME session as Window 3. ``` Go through your response sentence by sentence. Mark each one: [SPECIFIC] if it only applies to this incident, [GENERIC] if it could appear in any company's post-mortem unchanged. ``` Expected result: the model labels its own output. The specific parts are the four facts you gave it (migration, no backup, three hours, two customers) and the scaffolding around them. The generic parts are the majority of the corrective actions, lessons learned, and commitment language. The model becomes a witness against its own output. ## Beat 5: confabulation with and without prohibition Two variants of the same prompt. The data block is Stellex (see [`data.md`](./data.md)). First variant produces confabulated margins and growth rates. Second variant produces explicit statements of what is missing. Both variants can run in ChatGPT web. The video demo runs them through the API for cross-model consistency. See [`api_demo.py`](./api_demo.py). ### Variant 1: no prohibition ``` Stellex Corp Q3 FY2026 Financial Results Quarter ended September 30, 2026 Revenue by segment: - Total net sales: $2.847 billion - Cloud Platform: $1.234 billion - Enterprise Solutions: $0.876 billion - Data Services: $0.412 billion - Hardware: $0.325 billion Based on the data above, provide a financial analysis covering: 1. Revenue mix and segment trends 2. Estimated operating margins by segment 3. Year-over-year growth rates by segment 4. Key risks for the coming quarter ``` Expected result: the model invents operating margins (no margin data was provided) and year-over-year growth rates (no prior-period data was provided). Output includes authoritative language ("Cloud Platform likely operates at 65-70% gross margin") that has no grounding in the data block. ### Variant 2: prohibition added Same data block, same analytical asks. One paragraph added at the end. ``` Stellex Corp Q3 FY2026 Financial Results Quarter ended September 30, 2026 Revenue by segment: - Total net sales: $2.847 billion - Cloud Platform: $1.234 billion - Enterprise Solutions: $0.876 billion - Data Services: $0.412 billion - Hardware: $0.325 billion Based on the data above, provide a financial analysis. IMPORTANT: Use only data from the source above. For any metric not provided (margins, growth rates, operational data), explicitly state that it is not available rather than estimating. ``` Expected result: the model states which metrics are missing. Gemini lists them explicitly. GPT turns the gaps into risk analysis. The constraint does not just stop fabrication. It redirects the output toward better analytical structure. The line that does the work: ``` Use only data from the source above. For any metric not provided, explicitly state that it is not available rather than estimating. ``` This is the prohibition line. Insert it before the analytical ask, not after. Before the question works. After the question produces monitoring-style output where the model flags its own estimates but still provides them. The ordering matters.