Oswald's limits

Oswald is useful within bounds. This page names the bounds explicitly so the brand owner knows where Oswald’s drafts need closer human reading and where Oswald should not be used at all.

​

What Oswald is reliably weak at

​

Strategic decisions

Oswald has no opinion about the brand’s positioning, audience, or mission. It can summarise what is in the Strategy section; it can describe options; it cannot decide between them. A prompt like “should we move into the SMB market” returns options and trade-offs; the decision is not Oswald’s. A useful test: if the answer to the question would meaningfully change the brand record, the question is strategic. Decide it first, then ask Oswald to draft against the decision.

​

Asset selection

Oswald cannot pick the image. It can describe an image, brief a shoot, write a prompt for an image-generation tool. The actual selection of which photograph, which logo variant, which file to include is the brand owner’s. Oswald’s recommendations on this are weak; ignore them.

​

Fact-checking

Oswald is not a fact-checker. If an asset claims a statistic, a date, or a person’s title, Oswald will not verify it against external sources. The Voice and Typography checks Oswald runs during asset review are reliable; the factual claims need a human read.

​

Cultural reading

Oswald applies the brand’s documented rules. It does not intuit a cultural moment that should reshape the work. A campaign launching during a regional event, a press response to a current news cycle, a tone shift required by a sensitive moment: these need a human reading from someone close to the situation.

​

Original creative work

Oswald composes from material that exists in the brand record. Original ideas (the campaign concept, the unexpected typographic move, the photograph that defines the next quarter) come from the brand owner, the studio, or the team. Oswald is good at scaling and structuring an idea; it is not good at originating one.

​

Reading complex external material

Oswald reads the brand record reliably. Attachments to a session vary:

• Text-heavy PDFs. Reliable.
• Image-heavy PDFs. Reasonable for the images Oswald can interpret; less reliable for designed layouts.
• Interactive prototypes. Static exports work; live prototypes do not.
• Spreadsheets. Limited; better to paste relevant cells into the prompt.

When in doubt, paste the key content into the prompt rather than attaching the file.

​

What Oswald cannot do at all

Five hard limits, restated:

1. Apply changes without approval. Every change runs through brand-owner confirmation.
2. Read external systems. No email, no Figma, no Slack, no file system. The atlas is the world.
3. Make brand decisions unilaterally. The brand owner is the author; Oswald is the drafter.
4. Remember sessions beyond what is in the brand record’s history. Each session starts fresh; long-term memory lives in the record.
5. Generate images or video. Text only. Image and video work happens in other tools.

​

When the brand owner should override Oswald

Three patterns where overriding is the right call:

1. The brand record is wrong. Oswald drafts against the record. If the record contains an out-of-date or incorrect entry, Oswald will use it. Override and fix the record.
2. The voice is being applied too rigidly. Oswald respects the Voice section as a baseline. In moments of genuine humour, sorrow, or urgency, the brand voice can stretch. Override the draft; later, consider whether the Voice section needs a “when to stretch” block.
3. The draft is plausible but generic. Oswald can default to safe drafting on prompts that do not give it enough to work with. If the draft reads like brand-template prose, the brief was probably under-specified. Rewrite the brief, not the draft.

​

When not to use Oswald at all

Three cases where Oswald is the wrong tool:

• Crisis response. A live PR moment, a customer-safety incident, an internal communications crisis. Decide off-platform with humans only; record the outcome afterwards.
• Sensitive personnel matters. Anything HR-adjacent should not touch an AI tool, regardless of the AI’s privacy commitments. Use the right tool.
• Legally-bound writing. Drafting that needs to satisfy specific legal requirements (regulated industries, financial disclosures, healthcare claims). Counsel drafts; the atlas records the outcome.

​

The reliability heuristic

A useful heuristic for when to trust Oswald and when to verify carefully:

• Maintenance-scale tasks. Adding a value, clarifying a rule, drafting a small change. Oswald is reliable; light review is sufficient.
• Cross-section consistency. Drafts that touch multiple sections at once. Oswald is reliable on the consistency; verify the actual content.
• Strategic-scale tasks. Drafts that imply a brand decision. Oswald is unreliable on the decision; treat the draft as one option among many.

The pattern that produces the best outcomes is treating Oswald’s draft as the input to your thinking, not the output.

​

Related pages