The model is portable. The workflow is the moat.

Weekly Claw #21 — Host Cheat Sheet

Hard constraints

• Runtime: 45 minutes. Stop at 44:00, outro by 45:00.
• Recording: tech check before record; cold open starts the file.
• No abandoned segments: if Tool Fight is teased, run it.
• No live research: if a claim is uncertain, label it and move.

Episode title

The model is becoming portable. The workflow is becoming the moat.

Cold-open lines to hit

1. “Last week was the price shock. This week is the ownership shock.”
2. “If intelligence is cheap, the question becomes who owns the workflow.”
3. “The model is becoming portable. Context is becoming capital.”

Henry clip targets

Clip 1 — enterprise model switching

> A consumer can switch models and vibe-check the output. An insurance claims system cannot vibe-check whether 99% accuracy became 90%.

Clip 2 — open-weight risk

> Open-weight does not mean risk-free. You still need evals, scoped permissions, and runtime boundaries. Otherwise cheap intelligence becomes expensive cleanup.

Clip 3 — portability

> If your prompts, tools, evals, logs, and memory cannot move, you do not own your AI workflow. You are renting your own operations back from a vendor.

Andy moderation moves

Use these to prevent drift:

• “I’m going to compress that into one sentence…”
• “Give me the enterprise version of that in 30 seconds.”
• “What would change your mind?”
• “Is that evidence, inference, or vibe?”
• “Let’s not live-research it; call it launch-week claims and move.”

Segment clock

| Segment | Start | End | Kill switch |

|---|---:|---:|---|

| Cold open | 00:00 | 01:30 | No banter |

| What happened | 01:30 | 16:30 | Six stories; keep each tight after Sol |

| Debate | 16:30 | 23:30 | Force prediction + falsifier |

| Tool fight | 23:30 | 30:30 | End with Explorer / Worker / Verifier |

| OpenClaw / Signal outside | 30:30 | 36:30 | Tie 7.1 to control-plane thesis |

| Audience Q | 36:30 | 40:00 | One question only |

| Buffer / clips | 40:00 | 44:00 | Pull best audience answers |

| Close | 44:00 | 45:00 | Website, X, YouTube, goodbye |

Claims to say carefully

• Sol autonomy: public incident reports are not a statistical sample; pair them with OpenAI-documented risk language.
• Kimi K3: “reported / launch-week claims / independent validation pending.”
• Benchmarks: “claimed by launch-week summaries” unless verified before air.
• Quota changes: “community reporting says…” unless checked from official account.
• GPT-Red: official OpenAI source exists; quote the 6x fewer failures claim as OpenAI’s claim, not independent truth.
• Thinking Machines Inkling: open-weights/customization story; avoid pretending it beats frontier closed models.
• Sam/Demis regulation: both propose US-led standards/oversight; flag regulatory-capture risk.

Audience question

What is the one piece of your AI workflow you would most want to carry across models: prompts, memory, tools, evals, logs, or permissions?

Best closing line

> If you cannot move your workflow, you do not own it. Follow the excitement.

OpenClaw 7.1 line

> OpenClaw 7.1 is the episode thesis in product form: not another model leaderboard, but the control plane around sessions, tasks, cost, mobile, channels, recovery, and connected coding agents.