A four-person content team just shipped six campaigns last quarter. The previous quarter, with the same headcount, they shipped two. The difference was not the headcount or the budget. It was that they finally let the agent do the boring parts.
Content is the easiest place to start a real AI project. The work is structured (briefs, drafts, edits, assets, scheduling). The brand has a written voice. The asset library is finite. The cost of being wrong is a typo, not a wire transfer. If you are looking for a low-stakes operational win, this is it.
The pattern below is one we have shipped four times now, in slightly different shapes. The skeleton is the same. The fancy bit is not the model. It is the brand guide.
Why content is the easiest place to start
Three reasons. One: every step is reviewable before it goes live, so you do not need a perfect first draft. Two: the inputs are mostly text and images, both of which models handle well now. Three: the failure mode is "the asset gets edited or scrapped," not "we sent the wrong wire."
The other reason is that content teams are usually under-resourced relative to demand. Every founder we know wants more case studies, more launch posts, more video, more LinkedIn presence. The team cannot keep up. AI shows up here as the agent who does the boring parts so the humans can do the parts that need taste.
Skills, not workflows
The mental model that works for us is "skills, not workflows." A workflow is a fixed sequence of steps: take a brief, write a draft, generate an image, schedule. A skill is a capability the agent has, that you compose into different workflows depending on what you need.
The skills we equip a content agent with usually look like this:
- Read the brand guide. Tone, vocabulary, banned phrases, examples of "this is us" and "this is not us."
- Read the asset library. Logos, fonts, color tokens, photography style, past campaigns to reference.
- Write a draft. Long-form, short-form, social, email. Same skill, different prompt.
- Generate an image. Through whichever image API you trust. We currently like a small set of them for different jobs (illustration, photography, graphic). The agent picks the right one for the task.
- Self-edit against the brand guide. Run the draft past the brand rules. Flag the parts that drift.
You compose those skills into workflows on the fly. "Write me a launch post + a hero image + three social variants" is one workflow. "Repurpose this case study into a one-pager + LinkedIn carousel + email blast" is another. Same skills, different recipes.
The brand guide is the system prompt
The single biggest variable in how good the output looks is how good the brand guide is. Most brand guides are deck-shaped. They show fonts, colors, a tagline, a few "do not" examples. They are written for designers to glance at, not for an agent to follow.
We rewrite the brand guide as a structured document the agent can actually reason over. Voice rules with examples. Vocabulary lists, including the words we never use. A bank of phrases that sound like us. A bank of phrases that do not. Photography rules with sample shots labeled "yes" and "no." Tone gradients ("formal end" vs. "casual end" with examples of each).
Once the brand guide is structured, the agent's drafts stop drifting. The first draft is 70% of the way there, every time. The human's job becomes editorial review, not generation.
A brand guide written for a designer is a slide deck. A brand guide written for an agent is a system prompt. Same intent, very different shape.
Image APIs and the asset library
Image generation is the part of this that has changed the fastest. Two years ago we would not have shipped AI-generated imagery. Now we do, with care.
The pattern is to give the agent access to two or three image APIs, each with a job. One for illustration, one for photographic-style hero shots, one for product or graphic compositions. The agent picks based on the task. The brand guide includes prompt templates for each, so the look stays consistent across generations.
The asset library matters as much as the API. We index every existing asset (past campaigns, headshots, product shots, logos at different scales) so the agent reaches for an existing asset before generating a new one. About 60% of the assets in a typical run come from the library. Only the remainder are generated.
The agent that reaches for the existing asset first is the agent your brand team will trust. The one that generates from scratch every time will get switched off in a month.
A real run: idea to asset in 18 minutes
Here is what a typical run looks like, lightly anonymized. A brand manager opens a Slack thread. "We need a launch post for the new pricing page, plus a hero image and three social variants." They paste two competitor examples for tone reference.
The agent reads the request, pulls the brand guide and asset library, and produces five things in about eighteen minutes. The launch post in a Google Doc, the hero image as a PNG, three social variants in Figma-ready format, and a short summary of the choices it made (which images it picked from the library, which it generated, why).
The brand manager reviews. They approve two variants, ask for a tone shift on the third, and send the post to the editor for a final pass. Total human time on the work is roughly 45 minutes. The previous version of this workflow was four people, three days, half a Friday lost to image revisions.
What still needs a human
Editorial judgment. Strategy. The decision of what to write about in the first place. The agent is great at turning a brief into a draft. It is bad at deciding which brief to write.
The other thing that needs a human is the final read. We have not shipped a single piece of content without a human signoff, and we do not plan to. The agent does the boring 80%. The human does the 20% that has taste in it. That ratio holds.
If you have a content team that feels constantly behind, and a brand guide that lives as a slide deck, this is a four-week build. It is the cheapest visible operational win we ship.
We have built four versions of this for clients. If you want to see what one looks like for your stack, book a diagnostic.