How to Orchestrate Agents: The Demo Prompt
CxXchange SuccessHour, October 15, 2026 | Co-hosted with Jan Young
01 | The idea
If you need a wall socket added, you hire an electrician. One trade, one job. That's a normal AI chat: one question, one answer.
If you're remodeling a kitchen, you hire a general contractor. They hire the electrician, the plumber, and the drywall crew, and they hand you a finished kitchen. That's orchestration: one prompt, a whole job done, with the AI splitting the work across sub-agents and checking their work before it hands you the result.
02 | Write the scope of work: four lines
| # | Line | What to write |
|---|---|---|
| 01 | The job | The finished thing you want back, in one sentence. "A renewal proposal for every account," not "help me with renewals." |
| 02 | The why | Who it's for and what it's for, so the AI can make the small calls on its own instead of stopping to ask you. |
| 03 | The guardrails | What it can't touch, spend, send, or decide. Which data it may use, and what never goes to a customer. |
| 04 | Done | What finished looks like and how you'll check it, so it knows when to stop. |
The framework comes from Boris Cherny, creator of Claude Code at Anthropic, via Ben AI's video 7 New Rules for Prompting Claude 5 Models.
03 | The prompt
Four common customer success jobs, one prompt. The use cases came from Felicia Rella in the CxXchange Slack: renewal proposals, a quarterly ROI summary for leadership, CRM updates from call notes, and a usage gap analysis.
The job: Get our customer success team ready for this week. Produce four deliverables:
1. A renewal proposal for every account renewing within 120 days of October 15, 2026.
2. A one-page Q3 2026 ROI summary for our senior leadership team.
3. Proposed CRM field updates from this week's call notes.
4. Usage gaps for every account, each matched to the fix.
The why: So every CSM walks into this week with a recommendation and the evidence behind it, leadership sees what customers got for what they paid, the CRM matches what customers actually told us, and every account has a specific adoption play.
Guardrails:
- Split this into four sub-agents, one per deliverable, and run them in parallel. Before they start, tell me how you split the work.
- Use only the data below. Cite the file for every number.
- Renewals: use only the prices and pricing rules in product_catalog.md. No invented discounts.
- ROI: use the assumptions in README.md, show the formula behind every headline number, Q3 actuals only. List missing data as missing; never estimate it.
- CRM: propose changes only, never write to a CRM. Only change champion, renewal_risk, competitor_in_eval, expansion_opportunity, and next_step. Quote the sentence each change came from. If a note is ambiguous, list it as a question.
- Usage gaps: every gap needs evidence from feature_usage.csv or support_tickets.csv. Only recommend fixes listed in product_catalog.md.
- This is a draft for the team. Nothing goes to a customer.
Done: One briefing with four sections, one per deliverable. Then a short list of anything you could not verify. Then the three things the team should do first this week.
This prompt expects data pasted below it. The ready-to-run version, with all the fictional data already included, is the one-paste file below.
04 | Run it yourself
- Download the one-paste file. It's the prompt plus all six data files for Lumen Desk, a made-up help desk software company with ten customers. Every company, person, and number is fictional.
- Open Claude (the desktop app works best), paste the whole file, and send.
- Watch the first minute. Before any work starts, Claude should tell you how it split the job into four sub-agents.
- Give it about five minutes. You get back one briefing with four sections, a list of what it could not verify, and the three things to do first this week.
The full kit, with each data file and four single-job prompts you can run on ChatGPT, Gemini, or Copilot, is in the demo kit folder.
05 | Make it yours
- Paste the prompt into your AI and ask: "Help me rewrite this for my team, my data, and my weekly work."
- Swap in your own data only if your company's AI policy allows it.
- Keep the four lines. They're what keeps the AI on track.
- A person still reads the output before anything goes to a customer.
06 | Spoilers: what a good run catches
Run the demo before you read this.
- Renewals: five accounts fall inside the 120-day window. Quillstone renews February 28, just outside it, and should be left out.
- ROI: Copperline gets back about 10 cents for every dollar it pays. Northgate has no Q3 data, and that should be flagged, not estimated.
- CRM: Hollis Ridge's "might consolidate some tools" should come back as a question, not a change.
- Usage gaps: Copperline's answer bot handles 3% of tickets because it only has four help articles to learn from.
- A bonus catch: two accounts show auto-resolved tickets without owning the add-on that resolves them. That's a mistake in the made-up data, and the live run caught it.
07 | Where you can do this today
A side-by-side of Claude, ChatGPT, Gemini, and Copilot, with a source for every cell: labs.spearworksai.com/compare/agents