H3 Context Preview

Generating five seconds of video is not a cheap button. Queueing one to find out that your prompt was missing half the shot is the expensive way to learn something a text box could have told you. This is the text box.

What it does

It shows a bounded, redacted projection of the context report and the source prompt that produced it. The node's contract is presentation only - "preview must not add hidden prompt behavior" - so it doesn't re-render, tidy or extend anything. What you see is what the pipeline built.

"Redacted" is doing real work. The projection is content-free in the parts that matter: no file paths, no credentials, no provider responses. It's the same discipline the sidebar's Copy diagnostics button follows, and it means a screenshot of the preview is a safe-ish thing to post when you're asking for help. The "bounded" part is about size - the projection has a fixed ceiling, so a giant prompt doesn't produce a giant UI payload.

Inputs and outputs

One required input: report (H3_CONTEXT_REPORT), whose tooltip describes it as the "inspectable typed context report from Plan or Compiler." So you can preview the plan before rendering prompt text, or the compiled result after. Both are useful and the node doesn't care which you hand it.

Two outputs:

• prompt (H3_PROMPT_STRING) - the source prompt string, passed through.
• preview (H3_CONTEXT_PREVIEW) - the structured preview value. It's additive by design, so anything consuming it downstream is reading a projection, never the authoritative report.

The prompt output is what makes this node pleasant to use in a graph: it isn't a dead end. You can put the preview where you'd put the Compiler's prompt, and get the human-readable view and the wire in one node.

How you actually use it

In the sidebar, you already have this: the Context page shows the prompt, the plan behind it, and a validation report before anything is queued, and the five numbered stages (Intent/Mode, Media/Roles, Understand/Plan, Audit/Validate, Execute/Export) are the same pipeline exposed as tabs. What the canvas node gives you is the same view inside your own graph.

The workflow is boring and effective. Build the request, plan it, compile it, preview it, and read the diagnostics. Errors must be fixed; warnings are uncertainties you're choosing to accept. Fix, re-plan, preview again, and only then queue. If you're going to spend four minutes of GPU time on a clip, sixty seconds of reading is the highest-leverage thing available.

The preview is also where the audit trail starts. H3 Context Audit Override requires the exact report fingerprint from the preview - the one whose tooltip tells you stale edits fail - so a preview that's connected to your pipeline is how you get a fingerprint that matches what's actually on the wire.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-MiniMaxH3-Studio.git
# restart ComfyUI

The pack isn't in the Comfy Registry yet, so the clone is the install path. No Python dependencies declared, nothing fetched at install, prebuilt browser extension, Python 3.10+. This node needs no weights and no models; it's a view.

Two failure modes worth naming

The first: treating a passing validation as a quality verdict. It isn't, and this pack is unusually honest about that. Validation and the guide-readiness check cover the prompt - structure, hard constraints, whether the official prompt guide's expected information is present. Neither has any opinion about whether the resulting footage looks good. Everything in this pipeline is prompt hygiene; the aesthetics are still your problem.

The second: reading the preview once, changing the request, and reusing what you saw. The prompt is deterministic given the plan, and the plan depends on the request. Change anything upstream and the preview you're looking at is history. That's also why an Incomplete guide-readiness result is worth reading rather than dismissing - it usually names one specific missing thing, like sound being unspecified when you meant silent, and that fix belongs in the intent graph, not in the prompt text (prompt-engineering.md).

One last practical note: if the preview shows you a phrase you didn't write, don't hunt for the node that added it. The pipeline doesn't embellish. What's usually happening is that an upstream node's widget holds text you forgot you typed, or an intent graph is supplying structure you didn't remember connecting.