back

AI writes a draft in seconds. Checking it takes all afternoon.

Product DesignAIWorkflow DesignSaaSContent Systems

ContentOps

An AI writing tool built around the check, not the draft

Role

Solo Product Designer

Timeline

4 days

Scope

Research, UX, UI, coded prototype

Status

Self-initiated concept

What it is

A workspace for small content teams publishing blogs with AI. One person briefs and edits. Another decides what goes live.

The ContentOps dashboard on a desktop monitor: a sidebar for dashboard, jobs, library, releases, assets and analytics, a Quick create button, and Priya's view of what needs review and what is still running
One workspace from brief to live post

AI tools race to write faster. Writing was never the slow part. Checking was.

Today — the checks live in other tabsAI draftDocsSEO toolSlackCMSfixfixEvery check is another tab, and everyfix sends you back to the doc.ContentOps — the checks travel with the draftyou approveyou releaseResearchStrategyDraftCheckLiveSEO score 84Tone on brand98% originalOne job, one place. Nothing ships without a person.
The whole case study in one picture

The problem

I started where most AI writing tools do: how do we draft faster? Then I watched people use the drafts.

They rewrite

The AI's voice isn't the brand's. Whole paragraphs get redone.

They re-check

Facts, links and keywords, by hand, in a different tool for each.

They re-ask

One weak section means regenerating the whole post and losing the good parts.

Some said they spent longer fixing AI output than they would have spent writing. The problem isn’t speed. It’s trust.

So the product stopped being a generator. It became a pipeline — stages you can see, checks attached to the draft, and two moments where a person has to say yes.

My role

A self-initiated project, done alone in four days. No client and no brief — I set the problem as well as the answer.

What I owned

  • Research the problem through conversations and desk research
  • Map the two-role workflow, information architecture and flows
  • Design the interface, from dashboard to release gate
  • Build a working prototype with AI-assisted code

Two roles, one handoff

A small team splits the work in two. The design splits the same way.

Priya · Member

Growth & SEO lead

Does — briefs, approves the outline, edits, submits.

Needs — to know if the AI is still working, and to fix one section without starting over.

“I need to go from an idea to a finished preview fast.”

Arjun · Admin

Founder

Does — sets up the workspace, reviews, releases.

Needs — one report that says the post is safe, before it reaches the live site.

“I’m terrified of it posting something robotic or off-brand.”

Same app, two front doors. Switch role and the dashboard changes its question — from what should I fix? to what can I release?

Priya — what needs me, what’s still running
The member dashboard: Hello, Priya; counters for jobs assigned, drafts reviewed and review turnaround; a Needs review list with a draft-ready post and a post the admin sent back asking for a more casual tone; and an Active operations panel with two jobs still drafting and researching
Arjun — what’s ready to go live
The admin workspace overview: counters for active jobs, posts released this month and average time to release; a Ready for release list where each post shows who approved it and its SEO and originality scores, with Release on one and Resolve SEO on another; and a team activity feed
The working filesJourney maps, the full information architecture and the two-role flowShow

Priya’s map dips at one stage: review. That’s where she stops trusting the draft — and where most of the product ended up.

Priya's journey map across discovery, intake, waiting, review and iteration, with her mood dipping to critical at review and recovering to empowered once she can revise a single section
Priya — confidence drops at review
Arjun's journey map from configuring the workspace to releasing a post
Arjun — the fear is the last click
The full information architecture: workspace pages along the top, an intake modal splitting into a guided form or AI chat, then research, strategy, drafting, assets and quality checking, with failure paths looping to a stage-failure state and a release gate at the end
Information architecture — every stage can fail and retry
The two-role user flow: Priya submits a brief that runs through research, strategy, drafting, assets and checking to a review surface; rejected drafts go back to drafting; approved drafts reach a release gate only Arjun can pass
User flow — Priya’s lane hands off to Arjun’s at the gate

A post, start to finish

Six steps. Two of them are people saying yes.

01 · Brief

Fill in a form, or just talk.

Some days you know the keywords. Some days you only have a hunch. Both paths make the same brief.

Quick create as a guided form: topic, target audience, a tone-of-voice picker with thought leadership selected, and SEO keywords
Guided form
Quick create as a chat: the assistant asks what we are writing today with two starter cards, while an output preview on the right already shows a generating title, tags and an SEO health check
Chat, with a live preview

02 · Approve the outline

No drafting until the plan is right.

The AI proposes headings, audience and keywords first. Fixing a heading costs a second. Fixing 1,200 words costs an afternoon.

The strategy brief: a proposed outline with a main headline and section headings, each with a one-line intent, beside a target persona card, a keyword strategy with traffic and intent tags, and a projected SEO score of 84; Edit parameters and Start generation sit at the top
Start generation is the first yes

03 · Watch it work

Stages, not a spinner.

Research, strategy, drafting, assets, checking. Priya can leave and still know where her post is.

The jobs list: five jobs, three in progress, one completed and one blocked in red, each row with a stage tag such as Draft ready or Drafting, a progress bar and a due date
Every job names its stage — and a blocked one shows up in red

04 · Review with the evidence beside it

The SEO report and the AI’s reasoning sit next to the text.

Keyword counts, a visibility score, and a plain-language note on tone and logic. No other tab.

The draft at the checking stage: a five-step progress bar from research to checking above the article, a word count and history link, and on the right an SEO panel with 82 percent visibility, top keywords and a recommendation to add two internal links, above a dark AI reasoning card describing the tone and logical flow with a Refine content button
The progress bar from step 03 stays on top of every draft

05 · Revise one part

Highlight a sentence. Say what’s wrong. Only that changes.

The rest of the post keeps its edits. Version history is one click away if the new line is worse.

Revision mode: one highlighted sentence in the article, a revision popover asking how should I change this with a Regenerate button and a short why-this-pattern note, and Discard changes, Request revision and Approve and publish along the bottom
A revision is scoped to what you selected

06 · Release

Only the admin can publish, and only when the checks pass.

SEO above 80, no broken links, originality verified. A failed check greys out the button and says why.

Final review before release: drafting, optimization and release steps, three passed checks for SEO score above 80, no broken links and originality verified at 98 percent unique, and the destination, Editorial Hub blog
The second yes
Release management with one post blocked by a meta description that is too long, its Release now button disabled
Blocked — with the reason on the row

The decisions underneath it

Four calls, each one a trade. None of them is free.

01

Approve the plan before a word is written.

Most rework traces back to a wrong angle, not a wrong sentence. The outline is the cheapest place to catch it.

The cost — one extra step before anything appears. The tool feels slower to people who wanted a draft now.

02

The evidence sits beside the draft, not in another tool.

SEO, quality and the AI’s reasoning share the screen with the text they describe. You judge in one place.

The cost — a busier screen. The panels compete with the writing for attention.

03

Revise a part, never regenerate the whole.

Regenerating throws away every edit you liked. Scoped revision keeps them, so fixing one line never breaks three others.

The cost — a rewritten sentence can clash with the paragraph around it. The person has to read for seams.

04

Only one role can publish.

A release gate makes one person accountable for what goes live. That is what lets the founder stop worrying.

The cost — the admin becomes the bottleneck. If Arjun is away, nothing ships.

What changed

The first dashboard showed everything. Charts, counts and activity, all at the same volume. More to look at, nothing that said what to do next.

The shipped one answers two questions. What needs me? What is still running? Everything else moved to its own page.

First pass — everything, equally loudSix widgets. Zero next steps.Shipped — what needs you, what’s runningNeeds reviewRunningEvery item ends in a button.
Redrawn from memory — the first version wasn’t kept

The system

Warm, restrained, one loud colour. Orange means act — start, release, revise. Everything else stays quiet so the writing stays the focus.

The design system: Lexend and Plus Jakarta Sans type, a kinetic-orange primary with warm neutral surfaces, primary, secondary and destructive buttons, inputs, status pills for live, draft, failed and scheduled, data cards, and an empty state
Status pills carry the pipeline — live, draft, failed, scheduled

I designed it, then built it as a working prototype with AI-assisted code. A pipeline only makes sense once it moves.

Where it stands

A concept with a clickable build. Not yet tested with a real team.

The first test would time one post, brief to release, against the team’s current tools — and count the tabs. Next on the list: two people editing at once, and a direct push into real CMSs.

What I take from it

AI isn’t the product. The workflow around it is.

A model will always be a little wrong. The design job is making that wrongness easy to see, and cheap to fix.