Creative Agent — Plain-English Guide
Audience: content teams, marketing leads, reviewers. No technical background needed. Engineering version: technical.md. Back to the agents index.
In one line
You pick a topic; it writes the draft — under your brand guidelines, with your SEO keywords worked in, and it stops before publishing so a human decides.
What it produces today
| Output | State |
|---|---|
| Blog drafts | Complete — this is the mature path |
| Video scripts | Complete, including enforced hook formulas |
| Hooks | Opening lines generated against creator formulas |
| Images, GIFs, video | Not built. See gaps below |
If you're reading this to understand how the system works, read the blog path. The others follow the same shape.
How a blog gets made
You queue a news item. That's your entire involvement until review.
Generation runs in the background and stops at "ready". It never publishes on its own.
What happens in between is not one pass at writing — it's a draft followed by several correction rounds:
- Writes the draft using your brand context
- Cleans the markup and makes sure the source article is credited with a link
- Bolds your SEO keywords on first use
- Checks its own work — structure, length, keyword placement — and if something's wrong, rewrites once to fix it
- Runs a compliance check and can regenerate with that feedback
- Checks whether it reads as AI-written, and if so rewrites up to twice to sound more natural
- Scans for blocked phrases
- Saves as a draft for you to review
Steps 4 and 6 are the ones people don't expect. The agent grades itself and retries before you ever see the output.
What shapes the writing
The agent isn't working from the news article alone. It assembles:
- Your brand voice and your "avoid" list
- Your blog master prompt — see below
- Your compliance rules
- Your SEO keywords, short-tail and long-tail
- A competitor and trend brief
- Any competitor posts you attached to the queue item
- Your CMS address, so internal links point to the right place
The blog master prompt
This is the single biggest lever you have, and it lives on the Guidelines page.
Whatever you write there is inserted into every blog generation as a brand brief that outranks the system's built-in defaults for tone, structure and SEO. If your brief contradicts the generic guidance, your brief wins.
Two things it cannot override: compliance rules, and the technical format the system needs to save the draft.
Leave it blank and you get sensible built-in defaults. Fill it in and you're steering.
It applies to blogs only — deliberately. It won't leak into video scripts or image prompts.
Video scripts: hook formulas are enforced
For scripts, the creator's hook formula isn't a suggestion — the system applies it and checks the result.
There's also a cooldown: it tracks how recent scripts opened and steers away from repeating the same pattern, so a week of scripts doesn't all start identically.
Scripts keep full version history — regenerate, compare, restore, and patch individual scenes rather than rewriting whole.
Reviewing and refining
Drafts open in an editor. To change something, use the refine chat — describe what you want in up to 600 characters.
Refinements come back as proposals you Accept or Reject. Nothing is applied silently, and the proposal shows only the paragraphs that changed — old wording struck through, new wording under it — rather than handing you a fresh copy of the whole article to compare by eye. The full before/after is one click away when you want it.
If you edit the draft by hand while a proposal is still open, it is marked stale and can no longer be accepted: accepting replaces the whole body, which would quietly throw your edit away. Ask again and you get a rewrite of the current text.
This is also how compliance fixes arrive: "Ask AI to fix" on a compliance finding sends the flagged phrase and the rule behind it into this same chat, so the rewrite shows up as an ordinary proposal rather than something special you have to learn.
When something goes wrong
The system draws a firm line between two kinds of failure:
Structural problems stop the draft. If it comes out malformed or missing required parts after the repair pass, no draft is saved and it's marked for retry. Better nothing than something broken.
Quality problems produce a draft with warnings. SEO keywords that didn't fit, a humanization pass that had to be reverted, an unavailable checking service — all of these save the draft and tell you what happened.
| What you might see | What it means |
|---|---|
| "No SEO keywords matched this topic" | Your keyword inventory has nothing close to this topic |
| "Keywords assigned but not used" | They were selected but didn't fit naturally |
| "AI-detection check skipped" | The detection service was unavailable — the draft is fine, just unchecked |
| "Humanization failed; original draft retained" | The rewrite made things worse, so it kept the original |
The principle: you see every soft failure. The agent never quietly decides something was good enough.
Who can do what
Blog generation, editing and publishing are open to anyone who owns the tenant. There's no separate permission to grant.
The one admin-only thing is connecting the CMS itself — the credentials that say where drafts get sent. That's in Settings → CMS integration.
What it does not do
It doesn't publish. Every path stops at "ready". Sending to the CMS is a human click, and a high-severity compliance finding blocks it until someone acknowledges — see the Compliance Agent.
It doesn't make images or video. This is the largest gap against the roadmap, which calls for static creative, GIFs and video. Today the agent writes text.
It doesn't post to social. Scripts generate and land on the content calendar; publishing them is manual.
It doesn't do vernacular versions. Local-language output appears as a requirement in three places on the roadmap. Nothing does it yet.
It gives you one angle. The roadmap asks for different approaches producing different themes; today one prompt produces one take on a topic. Want a second angle, regenerate.
There's no creative library. Drafts live per tenant. Nothing indexes reusable assets, templates or past work across them.
Where this sits on the roadmap
This is deliverable ① and ② on the whiteboard's build list — creative and blog, end to end. Text is done. The build list has five more items behind it: social media automation, dashboards, ORM, the marketing OS architecture, and the creative repository.
Subagents
The blog path runs several of these in sequence on a single draft — which is why a draft you see has usually been through three passes, not one.
- Drafter — first pass from brand context, keywords and competitor brief
- Quality repairer — grades its own output on structure, length and keyword placement, then rewrites once
- Keyword selector — picks short-tail and long-tail per topic
- Humanizer — ZeroGPT detect, then up to two structure-preserving rewrites; reverts if quality regresses
- Refiner — the Accept/Reject chat, 600-character instructions
- Hook generator — candidate openings against the creator's formula
- Scene patcher — scene-level edits to a script without rewriting the whole
Deterministic, deliberately not subagents: hook-formula enforcement and the cooldown that stops a week of scripts opening identically.