Runway Agent custom preferences vs a Sume run instruction

Runway Agent's custom generation preferences are plain-language guidance such as Always prefer 1080p and 16:9, which Agent weighs when picking a model but does not treat as a strict rule. On a Sume Format run the instruction field is composed after the Format body, so where the two disagree the model follows what you asked for. One is a standing preference for a person's sessions, the other is a per-call override sent by your code.

Preference or override

Runway's examples

Runway's page gives three: Always prefer 1080p and 16:9, Use GPT Image 2 for images, and For long videos, prefer Seedance 2.5. It says short, direct instructions work best, and that if your request needs a capability a named model lacks, Agent may pick a different model that can complete the task.

What Sume tells the agent, in order

The agent receives a pointer at the Format, the attached package, your instruction (or the Format's default), a notice that the run is unattended, a pointer at your input file, and any attached images. Because the Format comes first as the how, your instruction can narrow or override it. Open the run's thread in Agents and the first message shows exactly this text.

curl -sS -X POST "https://api.sume.com/v1/formats/sume/sume-slideshow/runs" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: deck-77-v1" \
  -d '{
    "instruction": "Vertical 9:16. Keep every slide on screen for the same length.",
    "attachments": [
      { "type": "input_image", "image_url": "https://cdn.example.com/a.jpg" },
      { "type": "input_image", "image_url": "https://cdn.example.com/b.jpg" }
    ],
    "generation_spend_cap_usd": 10
  }'

Pick by who sets the rule

• A preference you want in every chat: Runway's Set as default.
• A rule that must hold for one automated call: put it in instruction and keep the data in input.
• A rule that must hold for every call: edit the Format, which bumps its version.