Talking to Oswald

Oswald responds to clear, scoped requests. This page collects the prompt patterns most brand owners settle into, with examples and notes on what to expect in each case.

​

The general principle

A useful Oswald prompt has three properties:

1. A clear ask. Draft, summarise, review, find, list, propose. The verb tells Oswald what kind of output is wanted.
2. A bounded scope. The section, the Horizon, the asset, the period. Oswald reads everything you give it; the bound focuses the work.
3. The intended use. Where the output will go, who will read it, what decision it serves. Oswald’s drafting adjusts when the use is named.

Most useful prompts fit in two or three sentences.

​

Patterns

​

Drafting a Horizon

Draft a Horizon for our Spring 2026 campaign. The campaign runs from March through May, leans into the photography direction we set in Q4, and uses Primary Blue at 70% as the dominant colour. Audience: all team members.

Returns: a structured Horizon draft with the campaign definition, visual system block referencing the existing Colour System, photography direction notes, tone notes, expected timeline, and a placeholder for assets.

​

Drafting an Update Request response

A team member is asking for a new tint of Primary Blue at 40%. Draft the response. Either approve and add the tint to the palette, or reject and explain why we keep tints to the approved set.

Returns: a draft response presenting both options with reasoning, structured for you to pick the path and apply.

​

Summarising a section

Summarise the Voice section in one paragraph for a freelance writer who will be joining for a six-week project.

Returns: a paragraph-length summary in the brand’s voice, focused on what a freelance writer will need to know quickly. Different from “summarise for an audit” or “summarise for a partner.”

​

Reviewing an asset

Review the attached PDF (our Q1 investor update) against the brand record. Flag any inconsistencies with Voice, Typography, or Colour System.

Returns: a structured review listing each inconsistency, the section it conflicts with, and a recommendation. Useful before publishing.

​

Finding gaps

Read the atlas and list the five most consequential gaps you see. Focus on places where a team member would reasonably struggle.

Returns: a ranked list of gaps with the affected section, a description of the gap, and a one-line suggestion for what to add.

​

Listing what changed

Summarise what has changed in the brand record over the last 30 days, grouped by section.

Returns: a structured summary of recent edits. Useful for catching up after time away.

​

What not to ask Oswald

Five patterns that consistently produce poor results:

1. Open-ended strategy questions. What should our positioning be? Oswald has no opinion about your brand’s positioning; it has summaries of what is in the record. Ask for options, not decisions.
2. Questions about facts outside the brand record. What is the latest typography trend? Oswald does not browse the web. Its world is your atlas.
3. Personal opinions. Do you think this colour is good? Oswald has no taste. It has rules, and it can apply them. Reframe as does this colour comply with our palette and accessibility pairings.
4. Comparisons across brands. How does our voice compare to [Competitor]? Oswald only sees your record. It cannot fetch a competitor’s brand book.
5. Image generation. Oswald does not generate images. It can describe an image, brief a shoot, write a prompt for an image-generation tool, but it does not produce visuals.

​

Iterating on a response

Three patterns for refining a draft:

• Ask for variants. Give me three options for the opening sentence.
• Tighten the scope. Rewrite this in 80 words instead of 200.
• Change the audience. Rewrite this for a vendor who has never seen the brand before.

Oswald iterates well. Most useful drafts come from two or three rounds rather than the first response.

​

Setting the response register

If you want Oswald to draft in a specific register (more formal, more direct, more conversational), include it in the prompt:

Draft the response in our standard voice, but slightly more direct than usual. Assume the team member has raised three similar requests this month and we need to be clear without being curt.

Oswald respects the brand’s Voice section as the baseline. Per-prompt register adjustments are interpreted within that baseline.

​

When Oswald asks you a question

Oswald asks a clarifying question when:

• The prompt is ambiguous.
• The brand record is silent on something the response requires.
• A decision is needed that the brand owner has not delegated.

Answer briefly. Oswald uses the answer to continue; the clarifying exchange is recorded in the session’s history.

​

Saving prompt patterns

On Guardian, recurring prompt patterns can be saved. Settings → AI → Oswald → Saved prompts. A saved prompt is a template the brand owner can re-run with parameters. Useful for monthly reviews, end-of-campaign summaries, and quarterly gap analyses.

​

Related pages