> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tryflowy.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Canvas Audit

> Have the assistant check your wiring, judge your results, and find what's missing.

export const DemoVideo = ({src, poster, caption, aspectRatio = "16 / 10", className = ""}) => {
  return <figure className={`not-prose my-6 w-full ${className}`}>
      <div className="overflow-hidden rounded-2xl border border-gray-950/10 bg-gray-100 dark:border-white/10 dark:bg-white/5" style={{
    aspectRatio
  }}>
        <video src={src} poster={poster} loop autoPlay muted playsInline className="h-full w-full object-cover" />
      </div>
      {caption && <figcaption className="mt-2 text-center text-sm text-gray-500 dark:text-gray-400">
          {caption}
        </figcaption>}
    </figure>;
};

The assistant isn't only a builder — it's a reviewer. Ask it to audit your canvas and it checks the graph for real problems, examines generated media pixel by pixel, and judges finished work against the brief.

<DemoVideo src="/videos/canvas/ai-audit.mp4" poster="/images/posters/canvas/ai-audit.jpg" caption="Asking the assistant to audit the canvas and fix the blockers it finds" />

## Check the wiring before you spend

Before generating, the assistant can run a deterministic check of the whole graph. It separates **errors** — invalid connections, inputs over a node's limit, nodes that can't generate, compose setups that aren't ready — from **advisories**, creative suggestions it applies only when they serve the brief. Errors are real blockers worth fixing before any credits go out; ask for this check after any big build.

> "Before we generate anything, check the graph for problems."

## Judge the results, not the prompts

A prompt says what you asked for; only the pixels say what you got. The assistant inspects an image's actual content when you ask it to verify a result, and for finished deliverables it can run a full critique: an adversarial judgment of the asset against its brief. When a result falls short, it comes back with a targeted fix, regenerates, and judges again.

> "Critique the hero image against the brief — does it actually read as premium?"

## Review layout and coverage

Audits aren't only technical. Ask the assistant to look at the canvas the way an art director would:

* **Layout** — "tidy this up": it auto-arranges the graph, groups related nodes, and frames sections so the board reads clearly.
* **Gaps** — "what's missing for a complete campaign?": it reads the canvas and reports what the deliverable still needs — a missing aspect ratio, an unscored video, a shot without continuity references.
* **Consistency** — "do these shots hold the same character?": it examines the media itself, not the prompts, and flags drift.

## Prompts to steal

* "Audit the canvas: blockers first, then suggestions."
* "Look at every generated image and tell me which ones miss the brief."
* "Check that the compose setup is ready before running the video."
* "What would you cut, and what's missing?"

<Tip>
  Audit at the checkpoints, not at the end: after the build (wiring), after the first generations (fidelity), and before assembly (coverage). Each pass is cheap next to regenerating a finished piece.
</Tip>