AI-Native Series · Part 1 of 4

Claude in the run-of-show: what works, what doesn't.

By Dr. Vincent W. Allen, DPS · May 2, 2026 · 5 min read · Part of AI-Native Broadcast Operations

A 23-page run-of-show lands at 6:14pm. Doors at 8:00am tomorrow. The client edited four sections since the last version. Here's the prompt that gets you to a usable cue list before dinner.

What Claude is actually good at, in this seat

Three jobs, all squarely in the model's strengths:

  1. Cue list extraction. A run-of-show document has segments, talent transitions, audio cues, video cues, lower-third triggers, presentation slide breaks, Q&A windows. Pulling those into a structured table is what the model does in 30 seconds and what an exhausted operator does in 90 minutes at 11pm.
  2. Version diffing. The client sent v3 at 6:14pm. You have v2. What changed? Claude tells you in a one-paragraph summary plus a section-by-section diff, in the format you specify.
  3. Conflict flagging. "Segment 4 ends at 9:42, segment 5 starts at 9:41. Audio source changes mid-talk — is the lavalier swap accounted for? Q&A is scoped for 15 minutes but the run-of-show only allocates 8." The model surfaces these in a numbered list.

What Claude is bad at, in this seat

Three jobs the model will hallucinate confidently, every time:

  1. Equipment-spec answers. "Will this cue work on a Blackmagic Constellation 8K?" The model will give you an authoritative-sounding wrong answer. Never trust the model on hardware specs. Cross-check against the manufacturer doc, every time.
  2. Live calling. "Should I cut to camera 2 or stay wide?" That's TD judgment under live pressure. The model has no situational awareness, no audio cue, no read on the talent's energy. It is not in the seat.
  3. Vendor-specific bug knowledge. "Does TriCaster TC2 still have the latency drift on the M/E1 bus when…?" The model's training data ages out of vendor firmware reality. Operators who've shipped on the equipment know. Models guess.

The prompt pattern I use

Here's the rough scaffold. Don't copy it verbatim — your client's run-of-show format is different. But the structure transfers.

You are reviewing a run-of-show document for a live broadcast.

Audience: institutional investors / corporate town hall / cultural broadcast.
Production stack: [Blackmagic Constellation, Crestron, Evertz, Resolume].
My role: Technical Director on the live desk.

Tasks, in this exact order:

1. Extract a structured cue list. Columns:
   segment_number, time_start, time_end, cue_type
   (audio/video/lower-third/transition), action, notes.

2. Flag any conflicts:
   - Segments that overlap in time
   - Audio source changes without a documented swap
   - Cue-type collisions (e.g., simultaneous LT triggers)
   - Q&A allocations that don't math out

3. Diff this version against [paste prior version].
   Summarize changes by section in plain English.

4. List the three things I should personally verify
   before doors. Do NOT include things you'd typically
   delegate to a comms producer or stage manager.

Do not invent specs. If a cell is ambiguous, say "ambiguous —
clarify with [role]."

A note on confidentiality

If you're working inside a Fortune 500 environment, the run-of-show is privileged information. Use the enterprise tenancy your client provides. Do not paste investor-day cue lists into a personal Claude account. The Intuit engagement I'm currently in operates inside a $500/month per-employee AI-development token budget — institutional access with appropriate data handling. That's the bar.

Bottom line

Claude in the run-of-show is a force-multiplier for the pre-show and immediately-pre-show window. It is not a TD. The judgment loop stays human. What changes is how prepared the human is when they sit down. Read the pillar essay for the full four-layer stack, or jump to GPT-4 for site reports for the post-show side of the workflow.

Run-of-show on your desk? Let's scope a Deep Dive.

Book →