How to review AI content before it’s published

Start with the work that will appear

Review the finished file in its intended placement: a correct draft can lose its font when rendered or use the wrong product name in an image. Keep the task, expected copy, required assets and brand guidance beside it. Campaign instructions apply to that work; they do not automatically change standing brand rules.

Compare specific elements with specific rules

Name the element, the rule it differs from and the correction:

Spelling can be compared directly; composition needs interpretation. If no rule or reference supports a finding, clarify the guidance. The voice-guide tutorial covers writing rules; keep original visual assets accessible too.

Request a brand Review

Rainbrand maintains context that agents use to make and check brand work. Its Review returns findings with reasons and available evidence pictures.

1. Supply a finished image or render a page, email or slide as a clear screenshot. Upload it if the agent cannot access the file.
2. Include expected wording for copy checks.
3. Select colour, typography, logo and copy as needed. The request takes one primary image; check required slides separately.

Review the attached finished composition against [brand]. Check colour, typography, logo and copy. Use the expected wording attached to this request. Wait for the Review, then show each verdict, its reason and any evidence pictures. Identify anything the supplied image or references did not let you check.

Check links, keyboard use and email-client rendering separately; a screenshot does not establish those behaviours.

Read the finding before acting on the verdict

Read each verdict with its reason and evidence:

For uncertain typography, inspect the source font before generating again. Investigate a not-applicable logo result if the task required one. An operation failure is not a verdict; a pass covers only the submitted input, references and requested checks.

Keep the evidence’s scope visible

For coffee roaster Friedhats, Claude made a phone page showing when to brew a delivered jar. Its first espresso-screen capture clipped the header motto and used Ubuntu Mono where the body guidance called for Archivo Black:

Claude made the motto a static line that fit and changed the body font:

The revised capture passed the four requested checks. The full Friedhats example keeps these matching captures and findings together.

Revise the work and review the result

Have your agent correct the editable source with its own tools, then check the revised output. Set a bounded loop:

Use the Review findings to revise this composition with your editing tools, then request another Review. Make up to two editing rounds. Stop when the applicable checks pass, an operation fails or a finding needs a decision you cannot resolve from the task and brand guidance. Show the final file, the changes and anything still open.

Two rounds is an example, not a product requirement. Choose a retry limit that fits the work and cost, and stop for missing files or unresolved facts. Inspect the whole revised composition: a line-break fix can crowd the logo.

Before recurring work, request a Brief, the guidance and assets selected for the task. Keep it with the output and this review record:

Finished file or rendering:
Intended placement and task:
Brand and guidance used:
Checks performed:
Findings and evidence:
Corrections and revised file:
Unresolved decisions:
Other publication checks, where relevant:

Keep any required approval with that same version; a brand verdict grants no publication authority. For recurring findings, correct the cause: guidance, missing files or production. If guidance needs changing, inspect and apply a brand edit, then request a fresh Brief.

Check what the brand review cannot establish

Keep these checks alongside brand findings:

Facts and claims. Ask the drafting agent for an evidence list, then open the sources:

List every factual claim in this finished draft, including features, numbers, prices, dates, comparisons and quotations. For each, identify the source and quote the supporting passage from the supplied material. Mark unsupported claims explicitly. Do not rewrite the draft.

Verify each quoted passage, following Anthropic's auditable-claims guidance. For advertising, FTC guidance calls for evidence of express and implied claims. Regulated or unfamiliar claims may need appropriate specialist review.

Rights and disclosures. Check permissions for the intended use of photos, fonts, logos and quotations, plus applicable channel/content disclosures.

Accessibility. Check the actual placement, including these common issues:

• Give informative images text alternatives that serve their purpose. Use the W3C’s alt decision tree to distinguish informative, functional and decorative images.
• Check text contrast, generally at least 4.5:1 for normal text and 3:1 for large text.
• Use real text instead of images of text where the technology can achieve the presentation, subject to the criterion’s exceptions.
• For video carrying prerecorded audio, check captions, including meaningful sounds as well as speech.

WebAIM’s Contrast Checker measures colour pairs, WAVE helps identify web accessibility issues, and Vale can apply written style rules to prose.

For the next piece, keep the output, task, brand reference and findings together. Give unresolved items a next action and retain any required publication decision.