Both the agent chat and direct node editing get the same canvas the same place. The choice is about leverage: which one gets you to the right output faster for this specific task? A rough rule, which holds about 90% of the time:Documentation Index
Fetch the complete documentation index at: https://docs.melius.com/llms.txt
Use this file to discover all available pages before exploring further.
- Use the agent for big-picture tasks, anything where you want the model selection and prompt writing done for you, and any bulk operation across many nodes.
- Use the node directly for fine-tuning, A/B comparison, and any change to a single specific output.
Use the agent when…
You’re starting from a brief, not a finished canvas
The agent is at its best when you give it a brief and a goal. “Build me five UGC video ads for this product.” It picks models, writes prompts, creates the right node types, and connects them up. You watch and steer rather than wire by hand.You want a bulk change across many nodes
“Update every video node on this canvas to use Seedance 2.0 instead of Veo 3.1.” “Re-generate every output node with the brand anchor connected as context.” “Add a CTA overlay to all the static ad variants in the winners group.” These are slow to do by hand. The agent handles them in one pass.You’re using a template you don’t know cold
Templates pre-configure a workflow (Image Style Analysis, Product Swapper, etc.), but they have parts you need to wire up. If you’re not sure which input goes where, just brief the agent: “Use the Image Style Analysis template with my reference images and output a style description I can use for new variants.” The agent will set it up correctly.You want clarifying questions before generation starts
The agent will ask follow-up questions when the brief is ambiguous — what aspect ratio, what audience, what tone. If you appreciate the safety net, use the agent in ask permission mode.You’re working asynchronously
Briefing from your phone, in a meeting, between tasks — the agent works without you watching. Set it to auto-run, come back when it’s done.Use the node directly when…
You’re fine-tuning a single output
You’ve got one image that’s 90% there. You don’t want to redo the canvas. Click into the node, edit the prompt — “keep everything, but change the camera angle to slightly higher” — and re-run. The original prompt is preserved as a previous version, so you can compare. This is the single most common edit pattern after the agent’s first pass.You’re A/B testing prompt variations
Two prompts, same model, see which lands better. Two nodes, two slightly different prompts, run both. The agent is overkill for this; doing it directly is faster.You’re learning what a model can do
When you’re new to a model — its style biases, its failure modes, what kinds of prompts it follows literally vs creatively — write a few prompts directly and inspect the results. You’ll build intuition faster than if the agent is between you and the model.You want exact control over the prompt
Sometimes you know exactly what you want to say to the model, and you don’t want the agent to “improve” your prompt. Open the node, write the prompt yourself, run.A worked example
Brief: “Generate 10 hero shots for a new product launch, in our brand style, sized for Instagram feed.” Agent pass (auto-run): the agent reads your brand anchor, runs an Image Style Analysis on your existing campaign references, generates 10 image nodes at 4x5 ratio with consistent style anchoring, hands you a canvas with results. Your pass (direct nodes): six of the ten are usable; four have issues. Edits, one node at a time:- Node 3 — product is slightly too small in frame. Edit prompt: add “product fills 40% of frame width.” Re-run.
- Node 5 — color drift toward green, should be warmer. Edit prompt: add “warm color grade, slight terracotta in the highlights.” Re-run.
- Node 7 — typography on the product label is illegible. Switch model from Nano Banana Pro to GPT Image 2 (High). Re-run.
- Node 9 — composition is fine but lighting is too flat. Edit prompt: add “strong directional side lighting from camera-left, golden hour quality.” Re-run.
A small but important habit
When you edit a node directly, the agent doesn’t automatically know about your edit. If you later ask the agent “do that thing again across the canvas,” it might revert to its original approach. Two ways to handle this:- Edit the brand anchor or a shared context node rather than a specific output. Changes there propagate to every node connected to it on the next run.
- Tell the agent what you changed. “I edited @node-7’s prompt to add ‘warm color grade.’ Apply that to the other warm-toned variants too.”
Push the agent further than you think it can go
Worth keeping in mind across both modes: most marketers undershoot what the agent can actually do. The instinct is to break a job into small briefs because that feels safe — describe the first piece, run, see the output, describe the next piece. That’s fine, but it leaves a lot on the table. The better instinct: brief the agent with the whole job, not just the next step. “Take these three reference ads, generate 10 angle variations across two products, resize each into 1x1 and 9x16, and run a brand-check pass against @brand-anchor.” If the agent can’t handle the full task, you’ll find out — and that’s useful information about where the current limit is. If it can, you’ve saved yourself an hour of orchestration. The cost of asking for too much is small (a re-brief). The cost of asking for too little, over and over, is your whole afternoon.Common pitfalls
- Re-prompting the agent to fix one image. Slower than just editing the node. Use the node.
- Editing the same node 20 times before realizing you’re solving the wrong problem. If you’re past your fifth tweak, step back and re-brief the agent with what you actually want. The agent has full canvas context and might unstick you.
- Treating the agent like ChatGPT. It’s not a chat. It builds canvases. Brief it like you’d brief a designer — concrete goal, references, constraints — not like you’re conversational.
- Undershooting the brief. Ask for more than you think the agent can do; let it tell you where the limit is.