---
title: "Generating, refining, and reviewing assets"
description: "How to generate image and video candidates from chat or a library, iterate on a candidate with feedback, hand work off to a designer, and review every run in the admin audit log."
---

# Generating, refining, and reviewing assets

This page is for anyone using [Assets](/docs/template-assets) day to day: asking for media, picking and refining a candidate, and — for admins — reviewing what was generated and by whom.

## Generate through chat or a library

Ask for what you want in plain language, from the home composer or from inside
a library's Generate controls: "three homepage hero options," "a square social
image in the launch-campaign style," "a video storyboard for the onboarding
flow." The message goes to the agent, which chooses references, a model, and
(when one matches) a preset, then returns candidates you can inspect, give
feedback on, and iterate on before anything is kept.

If you want the same brief across several destinations at once — three social
sizes, four slide backgrounds — ask for that directly ("give me a version for
each of these four slides") and Assets generates one candidate per destination
in a single pass instead of one request at a time.

## Iterate on a candidate

Once you have a candidate, keep going in the same conversation:

- **Give feedback** on a saved image and ask for a revision — "make it darker,"
  "move the subject to the right." Assets keeps the prior image as a reference
  and links the new candidate back to the source, so lineage is never lost.
- **Ask for a small, targeted edit** rather than a full regeneration — "remove
  the text in the corner," "fix the hand." This preserves everything else in
  the image.
- **Restyle while keeping the exact subject** — "put this product photo in our
  brand style" — when you want the library's look applied without changing
  what's actually in the shot.

## Logo accuracy in plain language

Image models are good at scenes and bad at logos — they smear gradients, drop
letters, and rearrange elements. Assets sidesteps this: when logo compositing
is on, the agent leaves a clean, empty area in the image instead of asking the
model to draw the logo, and the server stamps your library's actual uploaded
logo into that space afterward, pixel-perfect. The model never touches the
logo itself. This is a setting on the preset (or a one-off override for a
single generation) — see [Presets: design once, run everywhere](/docs/template-assets-presets#designers-build-one-preset-per-use-case) for how it's configured.

The same idea extends to a locked layout: a preset can carry a fixed brand
plate — a background image with an open region — so a webinar template's
header, framing, and logo survive every run untouched while only the open
region changes.

## Hand off to a designer

When someone else should keep working on a piece, ask the agent to hand it
off: "have a designer continue this." Assets creates a handoff session that
carries the brief, the chosen preset, the active candidate, your feedback so
far, and every run that led up to it. The designer can pick the work back up
in a fresh chat preloaded with that context, instead of replaying the whole
conversation or guessing at what was tried already.

<Diagram id="doc-block-assets5" title="Candidate generation loop" summary={"A brief produces several candidates; feedback and iteration narrow to a chosen asset that lands in the library. Every run — kept or not — is written to the audit log."}>

```html
<div class="diagram-genloop">
  <div class="diagram-node">
    Brief<br /><small class="diagram-muted">chat or Generate controls</small>
  </div>
  <div class="diagram-arrow diagram-muted" aria-hidden="true">&rarr;</div>
  <div class="diagram-box" data-rough>
    <span class="diagram-pill accent">N candidates</span
    ><small class="diagram-muted">refine · edit · restyle</small>
  </div>
  <div class="diagram-arrow diagram-muted" aria-hidden="true">&rarr;</div>
  <div class="diagram-col">
    <div class="diagram-node">
      Feedback<br /><small class="diagram-muted"
        >"make it darker", "try square"</small
      >
    </div>
    <div class="diagram-node">
      Chosen asset<br /><small class="diagram-muted">saved to library</small>
    </div>
  </div>
  <div class="diagram-arrow diagram-muted" aria-hidden="true">&rarr;</div>
  <div class="diagram-box" data-rough>
    <span class="diagram-pill">Audit Log</span
    ><small class="diagram-muted">every run, always recorded</small>
  </div>
</div>
```

```css
.diagram-genloop {
  display: flex;
  align-items: center;
  gap: 12px;
  flex-wrap: wrap;
}
.diagram-genloop .diagram-col {
  display: flex;
  flex-direction: column;
  gap: 10px;
}
.diagram-genloop .diagram-box {
  display: flex;
  flex-direction: column;
  gap: 4px;
}
.diagram-genloop .diagram-arrow {
  font-size: 20px;
  line-height: 1;
}
```

</Diagram>

## The generation audit log

Every generation — kept, discarded, or failed — is recorded automatically:
the prompt, the compiled prompt actually sent to the model, the model and
aspect ratio used, the references it drew on, who ran it and from where (chat,
the UI, or another app over A2A), and every asset it produced.

Org admins (or, in single-user local mode, anyone reviewing their own runs) see
an **Audit Log** link in the sidebar. It opens a filterable list — by date
range, owner, library, model, status, source, or a prompt search — with a
detail view per run showing its full references, generated children, and
(for a refinement) the parent run it came from. An **Export CSV** button
bulk-exports the filtered runs to a downloadable file, useful for a design
team reviewing quality by model, app, or prompt pattern across everything
generated.

<WireframeBlock id="doc-block-assets6">
  <Screen
    surface="desktop"
    html={
      "<div style='display:flex;flex-direction:column;gap:14px;padding:18px;min-height:520px;box-sizing:border-box'><div style='display:flex;align-items:center;gap:10px'><h1 style='margin:0'>Audit Log</h1><span class='wf-pill accent'>Org admin</span><div style='flex:1'></div><button>Export CSV</button></div><div class='wf-card' style='display:flex;gap:8px;flex-wrap:wrap'><span class='wf-pill'>Date range</span><span class='wf-pill'>Owner</span><span class='wf-pill'>Model</span><span class='wf-pill'>Status</span><span class='wf-pill'>Source</span><span class='wf-pill'>Prompt search</span></div><div class='wf-card' style='display:flex;flex-direction:column;gap:8px;flex:1'><div style='display:grid;grid-template-columns:2fr 1fr 1fr 1fr 1fr;gap:8px;border-bottom:1px solid var(--wf-line);padding-bottom:6px'><span class='wf-muted'>Prompt</span><span class='wf-muted'>Library</span><span class='wf-muted'>Model</span><span class='wf-muted'>Source</span><span class='wf-muted'>Status</span></div><div style='display:grid;grid-template-columns:2fr 1fr 1fr 1fr 1fr;gap:8px'><span>Three blog hero options</span><span>Acme Brand</span><span>gpt-image-2</span><span class='wf-pill'>chat</span><span class='wf-pill accent'>completed</span></div><div style='display:grid;grid-template-columns:2fr 1fr 1fr 1fr 1fr;gap:8px'><span>Webinar LinkedIn ad</span><span>Acme Brand</span><span>gpt-image-2</span><span class='wf-pill'>a2a · slides</span><span class='wf-pill accent'>completed</span></div><div style='display:grid;grid-template-columns:2fr 1fr 1fr 1fr 1fr;gap:8px'><span>Product cutout render</span><span>Launch Campaign</span><span>gpt-image-1</span><span class='wf-pill'>ui</span><span class='wf-pill warn'>failed</span></div></div></div>"
    }
  />
</WireframeBlock>

## What's next

- [**Presets: design once, run everywhere**](/docs/template-assets-presets) — build the recipe once so this workflow needs no settings
- [**Using Assets from other apps and agents**](/docs/template-assets-integrations) — the embedded picker, A2A, and Creative Context reuse
- [**Assets**](/docs/template-assets) — back to the overview