The article workbench: seven stages, and a person can take any of them

The problem behind it

The articles were the weakest format. The human read averaged about 4 for articles. Many were research memos. Some ran past 3,000 words and repeated one idea with a caveat in every paragraph. The best article scored 8 to 9 and did one thing: it argued a single surprising thesis and tied it back to the business.

Underneath that, two older problems. In October a Jio telecom article cost Rs 180 and was unusable, because every check ran after the writing. And the engine that ran articles had no place for a person to step in. Mid-stage, a human could only wait.

I wanted something I had seen work when I wrote with Claude in chat: research, angle, outline, draft, images, then checks, with me able to change my mind at any step.

How it works

The workbench lives at one page per article, opened from a Market Finder story card ("Write article from this card"). Seven stages, each visible, editable and takeable:

1. Research pack. The facts the article may use, each with an id (F1, F2), a source and a check. Verified client facts come first. Three web researchers fill gaps. Market Finder numbers for that market are added by code. You can add a fact by hand, or tick "my own experience", which also saves it to the client's verified facts.
2. Angle. One claim a reader could argue with, why this client may say it (with fact ids), the story, and two alternatives. It stops and says why if there is no proof point from the client's own experience.
3. Outline. Five to nine sections, each with its point and fact ids, internal links and next step.
4. Draft. Three writers, best one kept. Every fact sentence carries its id. Ids are stripped when the article is approved.
5. Images and cards. The market card is drawn by code from the Market Finder numbers.
6. Checker. Every claim against its fact id, the never-say list, leaked notes, worksheet-style output, and rules for the kind of article (Discovery, Trust, Conversion, Retention or Personal).
7. Human review. Who wrote it, how it was made, why it helps the reader, who reviews. Approve is blocked while a claim is unsupported.

Every stage has: redo with the AI (with a note), "I'll take this", hand back, "stop after this stage", and its own editor. Editing a stage marks later stages out of date. Nothing is redone until you press Redo. The estimate for the whole article is about Rs 45, with each stage's cost beside it. Nothing stops to ask.

The decisions

Stages live outside the work engine (ADR-445). The old engine runs steps in order with no person in the middle, and it would have started writing the old way on its own. So article stages have their own tables, and a story-card article is a human-mode task that the old flow never starts. The stage rules are plain code. A second person saving gets a conflict, not an overwrite.

Reviewing the plan caught this. A separate Opus critique of the next spec, which wanted to move article stages into the engine, flagged it as going against ADR-445. The workbench already had "I'll take this", "Hand back" and "write by hand". The fix was to open the existing workbench in person mode, not to rebuild.

One research machine. Two specs described the same research pack, citation checker and fact shape. A reviewer found it before build. It is built once.

Five purposes. Content is planned as Discovery, Trust, Conversion, Retention or Personal. I set rough limits: personal at most 5%, case studies 25%.

What went wrong

The Jio case: the research gate was missing, internal notes like "I have not opened any source" were written into the article body, and the checks ran four times. The workbench exists because of that.

Another miss: articles can double. Campaign articles, weekly rules and calendar slots run side by side. The help article says to use one source per client. That is advice, not a fix.

What's next

Verify live that one Mukika article runs end to end from a story card. The hand-over pattern could be copied to other formats later; that is a future wave, not built. I have no recorded cost per finished article from live use yet.

Numbers at a glance