Skip to main content
On Guardian, Oswald can draft a response to an Update Request. The brand owner reads the team member’s proposal, asks Oswald to draft a change against the current brand record, reviews Oswald’s draft, edits if needed, and approves. The decision stays with the brand owner; Oswald does the routine drafting.
Update Requests with Oswald is Guardian-tier only and currently in pre-general-availability beta. The feature is rolling out through 2026. If you do not see Draft with Oswald in your request panel yet, it has not yet been enabled for your account.

​
What Oswald does, exactly

Given an Update Request, Oswald can:
  1. Read the current state of the affected section. Oswald has access to the entire brand record.
  2. Read the team member’s proposal and reason. Including any attached material that Oswald can parse (text, links).
  3. Draft a proposed change. A specific, applicable change to the section: the new block, the new value, the new “don’t” example.
  4. Explain the reasoning. A short note on why Oswald drafted the change this way, including any uncertainties.
  5. Surface ambiguities. If the request is unclear, Oswald asks a clarifying question of the brand owner before drafting.
Oswald does not apply the change. Oswald drafts and presents; the brand owner decides.

​
The tool-use loop

Oswald’s drafting runs as a tool-use loop: a sequence of steps where Oswald reads the relevant material, decides what to draft, and presents the result. Each step is visible to the brand owner. The loop has explicit confirmation points:
  1. Read confirmation. Oswald confirms what it read: the section, the related sections, the request.
  2. Draft. Oswald produces the draft.
  3. Brand-owner approval. The brand owner approves, edits, or rejects the draft.
  4. Apply confirmation. On approval, Oswald applies the change. The brand owner is asked to confirm before the apply runs.
The brand owner sees the loop in the request panel and can intervene at any step.

​
Using Oswald on a request

1

Open the request

From the Inbox, click the request you want Oswald to draft.
2

Click Draft with Oswald

Oswald begins reading. The brand owner sees Oswald’s progress: “Reading Colour System current state… Reading related sections… Drafting proposal.”
3

Review the draft

Oswald produces a draft change, presented alongside the current section state for comparison. Oswald’s reasoning is shown below the draft.Read the draft carefully. Oswald’s drafts are usually close; they are not always right.
4

Edit if needed

Click Edit draft to modify Oswald’s proposal. The edit happens inline; Oswald notes the edit for future drafts.
5

Approve and apply

Click Apply. Oswald asks for a final confirmation. Confirming runs the apply. The change merges into the brand record, attributed to the brand owner with a note that Oswald drafted.
The request is resolved. The team member is notified. The brand record records the change, your approval, and Oswald’s drafting role.

​
What Oswald does well

Oswald is reliably good at:
  • Maintenance-scale edits. Adding a value, clarifying a rule, updating a list, adding a “don’t” example.
  • Cross-section consistency. Drafts that reference other sections correctly without re-stating them. A new colour value points at the Colour System; it does not invent a new system.
  • Voice alignment. Drafts in the brand’s voice, with banned vocabulary excluded.
  • Surfacing ambiguity. When the request is unclear, Oswald asks a clarifying question rather than guessing.

​
What Oswald does not do well

Oswald is reliably weak at:
  • Strategic decisions. A request that requires a strategic call (“should we widen the palette”) will get a plausible draft from Oswald, but the decision is not Oswald’s to make.
  • Asset selection. Oswald cannot pick images or files; it can recommend what kind, but the choice is the brand owner’s.
  • Original creative work. Oswald composes from material that exists. Original ideas come from the brand owner or the team.
  • Reading external material reliably. Attachments Oswald cannot parse (a PDF brief, a Figma file) limit its drafting.

​
Auditing Oswald’s work

Every Update Request handled with Oswald carries an audit trail:
  • The team member’s original request.
  • Oswald’s read confirmation.
  • Oswald’s draft, with reasoning.
  • Any edits the brand owner made.
  • The final applied change.
The audit is visible in the request’s history and in the brand record’s history.

​
Disabling Oswald per-request

If you prefer to handle a specific request without Oswald, simply do not click Draft with Oswald. The standard flow remains available. Oswald never drafts unprompted; the brand owner always initiates.

​
Disabling Oswald globally

Open Settings → AI → Oswald. Toggle Available on Update Requests off. Oswald remains available for Build-with-Oswald and other uses; the request panel no longer offers Oswald drafting.

​
Privacy and the AI providers

Oswald uses Anthropic Claude as its underlying model. Request content and brand record content are sent to the provider only for the drafting; nothing is retained beyond the session, and the API tier in use is the no-training tier. The full disclosure is in AI Usage & Disclosure Policy.

What Oswald does

The full Oswald reference.

Reviewing and approving

The non-Oswald flow.

Oswald deep dive

How Oswald works under the hood.