Content Agent: From Campaign Idea to Grounded First Draft

A good model can still produce a bad first draft.

Usually, some of the problems happen before generation. The campaign brief is in one place, product knowledge in another, brand rules somewhere else, and previous editorial feedback may only exist in someone's head. The person creating the asset has to remember what matters and translate it into a prompt.

Prompt-writing skill ends up influencing content quality more than it should.

We've been working on that problem in Magi's Content Agent: giving every assignment a more consistent base of context and putting human judgment into the workflow before the draft is written.

What Inputs Does Content Agent Use to Create a Draft?

Content Prompt, Campaign, and Keywords

Start by defining what the draft needs to achieve, who it is for, and the campaign it belongs to. If you’re creating an SEO-optimized asset, you can select one primary plus up to five secondary keywords.

At each stage, you can choose the AI model before generating the next artifact. Default (OpenAI) is recommended for most work. Choose Opus 5 when the assignment needs the most capable reasoning, Sonnet 5 for a balanced and efficient option, or Opus 4.8 when you want a more proactive, thorough pass.

Workspace Knowledge and External Files

Attach up to three relevant external files and three workspace knowledge assets from Magi. Workspace knowledge assets include connected product docs, campaign briefs, sales-call recordings, research, and notes from read-only Slack, GitHub, Notion, Google Drive, and Fathom connections.

Content Agent works from those selected sources. If you do not attach any sources, Content Agent will search the knowledge base and ideas for them instead.

Content Formats and Templates

Choose whether the asset should be a LinkedIn post, blog post, landing page, email, or one-pager. Each format has a different job. A blog post educates, landing pages focus on capturing specific leads, and one-pagers explain product features or capabilities in detail.

For each format, you can choose the specific template that you’d like to use for the asset. For instance, here are a few templates that Magi provides for blog posts:

Content Agent uses the selected format and template to shape the brief and outline accordingly.

Voice, Tone, and Editorial Boundaries

Set the voice and tone for the assignment, and whether Content Agent can use competitor material in its research and output. If you are using Magi for the first time, you can set up your voice and tone profiles in Brand Ontology.

You can also add “Skills“ to the prompt for specific tasks such as humanizing articles.

Catch a weak direction before it becomes a full draft

More context helps the agent understand the company. Choosing the right argument is a different problem.

An experienced writer knows when an angle feels obvious, when a section isn't earning its place, or when language is unlikely to survive review. Much of that judgment comes from accumulated editorial experience.

Magi's prompt systems encode editorial guidance around how content is written, including the patterns and language the system should avoid. The workflow then gives the marketer places to apply their own judgment.

Each assignment moves through brief → outline → draft.

The brief establishes the assignment. The outline makes the argument visible before the full piece exists. You can change the angle, reorder sections, remove an unnecessary point, add missing context, or work through a section with Copilot.

This is one reason we keep the outline as a distinct stage. Finding the wrong argument in an outline is much cheaper than finding it after the draft has been written.

Next, the outline turns the brief into an editable content plan. Each section appears as a card. You can change their order, add new cards, delete them, or ask Copilot, an AI assistant built around Magi’s custom editor, to help you improve the outline. You can also edit specific cards and the content within a card using Copilot. To manually edit a card, click inside a card and start typing. The outline lets you test the logic, emphasis, and flow before creating the first draft.

Finally, the approved outline turns into a working on-brand first version. Content Agent can also generate visuals in your draft through models like gpt-image-2, using BrandOS guidance to keep imagery aligned with your brand’s style and direction.

How does Content Agent Ground Claims With Citations?

Content Agent creates the first draft in Magi's content editor and grounds factual claims as it writes, linking them to public URLs, uploaded files, and internal knowledge sources. If it cannot ground a claim, it omits it.

In the editor, you can inspect the supporting source beside the claim, then keep, revise, or remove the language before publishing. When the draft needs a controlled edit, Copilot can help you refine the active document without rerunning the workflow.

How Do Content Agent and Copilot Work Together?

Content Agent runs the production flow from brief to draft. Copilot helps you refine the work in progress at each stage.

Guided and Autopilot are now part of the same workflow

Some assignments need that checkpoint. Others already have an approved direction and can move faster.

Content Agent now has Guided and Autopilot modes.

Guided pauses between the brief, outline, and draft for review. Autopilot can progress through all three automatically while streaming each stage. The Stop control remains available throughout an Autopilot run.

Autopilot shipped across Magi's frontend and AI service in August. It is controlled by the CONTENT_AUTOPILOT feature flag, so availability can vary by workspace.

The same release added run cancellation and resumable streaming. An active generation can be stopped while retaining partial output. If streaming is interrupted by a disconnect, the session can recover without duplicating chunks or leaving the draft stuck in a streaming state.

This gives teams a practical choice. Review the direction when the assignment calls for it. Let an established direction run when it doesn't.

What Is Live in Content Agent Today?

Available Now

• Create LinkedIn posts, blog posts, landing pages, emails, and one-pagers.

• Ground drafts in selected files, knowledge sources, campaign context, and BrandOS.

• Add inline citations to supported factual claims.

• Generate BrandOS-directed visuals alongside the copy, so the first draft includes a more coherent visual direction.

• Set competitor-use rules and work from a chosen template.

• Restore earlier draft versions as the story changes.

• Collaborate with teammates in the draft.

Current Boundaries

• Collaboration is currently limited to the draft, not the brief or outline.

• Separate outline versioning and a formal brief-review gate are not live yet.

• Citation behavior can vary by export or publishing destination.

What Can Marketers Build With Content Agent?

• Product launches: Create blogs, emails, landing pages, LinkedIn posts, and one-pagers for launch campaigns.

• LinkedIn content: Turn a founder perspective, product update, or customer insight into a first draft for LinkedIn.

• SEO content: Build source-backed blog drafts around a specific topic and selected primary and secondary keywords, then structure them for search and answer engines.

• BOFU pages: Build comparison or solution pages based on relevant product and competitor context.

• Sales enablement: Turn approved product evidence, buyer pain points, customer proof, and differentiators into a one-pager that sales and marketing can refine together.

How Does Content Agent Compare?

Jasper offers strong brand and project controls. ChatGPT and Claude offer flexible general-purpose assistance. Content Agent gives marketers a structured path to plan, ground, and draft a campaign asset before they start editing it.

Recent Content Agent updates

The current product now includes:

• Guided and Autopilot generation, with Autopilot availability controlled by workspace configuration.

• Stop, cancellation, and resumable streaming for longer Content Agent runs.

• Read-only Notion knowledge restricted to workspace-configured roots and their descendants.

• More reliable Idea and Knowledge selection, including clearing stale selections.

• Model selection within the Content Agent workflow.

• More resilient socket sessions, including stale-session eviction and a longer AI-service connection timeout to reduce duplicate streaming during reconnects and deployments.

These build on the existing Content Agent workflow for LinkedIn posts, blogs, landing pages, emails, and one-pagers, with BrandOS, campaign context, knowledge, templates, citations, and visual generation available as inputs to the work.

What we're working on next

There's a larger Content Agent update currently in development.

External references will allow the agent to ingest URLs, PDFs, images, and knowledge assets as material for an assignment. The work also includes an explicit option for cases where the user wants to reproduce elements of a reference's copy or layout.

We're also developing an intermediate HTML draft mode that converts deterministically into Tiptap. This gives the generation process a structured representation of the draft before it reaches the editor.

Recent chat context is being added as another input to generation, allowing relevant decisions from the conversation around an assignment to inform the resulting work.

And outline and draft generation are being brought into chat, so those stages can be triggered within the conversation rather than treated as separate interactions.

These capabilities are currently in development and aren't part of the live product yet.

A better first draft starts before the first sentence

We're measuring Content Agent against a more useful outcome than how quickly a model can fill a page.

We want teams spending less time gathering context and correcting avoidable mistakes. We want fewer revision cycles between an initial idea and something that can be approved. And we want the expertise surrounding the work—brand knowledge, editorial rules, evidence, and human decisions—to be available when the next assignment begins.

Content Agent is getting closer to that with each release.

New to Magi? Book a demo to see a launch assignment become a multi-asset first draft.

Already on Magi? Open Content Agent from the content workspace and start with a campaign or content assignment.

Frequently Asked Questions