| Format | Screen recording + voiceover (Loom, OBS, or native screen recorder) |
| Target length | 10–15 minutes. Aim for 12. |
| What's on screen | The agent conversation in ChatGPT or Claude. Full screen, no distractions. |
| Audio | Kathryn's voice narrating what's happening. No background music. No intro jingle. |
| Editing | Minimal. Cut wait times while the AI generates (or speed up). No transitions, no graphics overlays, no text animations. |
| Input data | Use an anonymized meeting list. Can be based on the case study firm or a composite. Must look realistic — real meeting names, real frequencies, real team roles. |
| Tone | Direct, conversational, calm. You're showing a colleague how this works, not presenting at a conference. Same voice as the brand. |
Create the ChatGPT Project (or Claude Project) with the system prompt from sync-tax-recovery-agent-v1.md. Run through it once before recording to make sure the output is clean. Note any spots where the agent asks a question or pauses — you'll want to have your responses ready so the recording flows.
Have an anonymized meeting list ready to paste. Suggested structure (based on the case study but genericized):
"The team uses Slack and Google Docs." (Keep it common — most of the ICP uses these or close equivalents.)
What matters here: The buyer sees what "good input" looks like. They see it's not complicated — just list the meetings with basic details. Lower the friction for when they do it themselves.
Tip: Paste the whole block at once. Don't type meeting by meeting — that's slow and doesn't reflect how people actually use it.
Pause on the team meeting classification. This is the "aha" moment for the buyer. Most practice owners think their team meeting is collaborative because people are in the room together. The classification reveals that the first 30 minutes are broadcast disguised as collaboration.
Don't comment on every meeting. Highlight the team meeting (the main event), one pure IT meeting (the check-in), and one pure CW meeting (the monthly planning). The buyer gets the idea from three examples.
Do not react with hype. Don't say "wow" or "can you believe it?" Read the number. Reference the benchmark briefly — "this is consistent with what we see in firms this size." Move on. The math does the work. Understatement is more credible than excitement.
This is a 60-90 second beat. Quick. The number lands and you move to the fix.
This is the longest and most important beat. The buyer needs to see the complete system being built — not a summary, the actual output.
As the memo structure appears:
As the AI project instructions appear:
Scroll through the full output. Let the buyer see the volume of what was produced. Don't rush past it. This is where $27 starts to feel like a steal — they're looking at output that would take a consultant hours to build.
Tip: If you need to cut time anywhere, don't cut this beat. The CEO Memo build is the product in action. Everything else supports this moment.
Read one line from the kill script to show the tone. Point out the resistance-handling language — the part that frames it as an upgrade, not a cut.
Close with the deployment timeline:
No pitch for Advisory OS in the video. The bridge lives inside the agent (Phase 5), not in the walkthrough. End on the deployment action, not a sales transition.
If the agent takes 15–30 seconds to produce output, speed those segments up in editing (2x or 3x) or cut to the result. The buyer doesn't need to watch text stream in real time. They need to see the input and the output.
The "say" notes above are talking points, not a teleprompter. The video should sound like you're showing a colleague something you've already used, not reading a script. Talk to the screen. React to the output naturally.
This doesn't need to be polished. It needs to be real. A clean one-take screen recording with natural narration is more credible than a produced video with cuts and transitions. If you stumble on a word, keep going — it sounds human.
Set your browser to 125% or 150% zoom so the agent text is readable at video resolution. Small text in a screen recording is the #1 reason these videos don't work.
Only the agent tab should be visible. No notifications, no bookmarks bar, no other tabs showing. Clean screen.
Use a decent microphone — AirPods are fine, laptop mic is not. The voiceover is half the value. If the audio is bad, nothing else matters.
| ☐ | Agent created in ChatGPT or Claude with system prompt loaded |
| ☐ | Agent tested — ran through once, output is clean |
| ☐ | Anonymized meeting list prepared (5 meetings, realistic names and details) |
| ☐ | Rates ready ($275 owner, $130 manager, $85 staff, $55 admin) |
| ☐ | Browser zoomed to 125–150% |
| ☐ | Other tabs closed, notifications off |
| ☐ | Screen recording tool ready (Loom, OBS, native) |
| ☐ | Microphone tested — clear audio, no echo |
| ☐ | Beat notes visible on second screen or printed (not on the recording screen) |