Judgement as infrastructure, part one: most editorial feedback is used once

Sandhya Simhan, who runs content marketing at OpenAI, shared something this week that looked surprisingly familiar.

Her team has started a content marketing repository. It contains a growing set of editorial agents built around recurring feedback, examples of strong work and standards for different formats. The aim is not to remove the editorial team. It is to help product marketers, engineers and other collaborators get a stronger first pass before a draft reaches them.

Her suggested process is practical: examine editorial comments and discussions, find the feedback that keeps recurring, gather strong examples, turn the patterns into focused agents or skills, then continue updating them as the team reviews real work.

This is interesting to me because it is very close to how I have been building Editorial Intelligence OS.

Not identical. Her system is a team-wide editorial quality layer inside one organisation. Mine is a more personal combination of knowledge system, publishing infrastructure and thinking environment.

But the underlying principle is the same:

Editorial judgement does not have to disappear after the edit. It can become infrastructure for the next piece of work.

Most editorial feedback is used once

A large amount of editorial judgement currently lives in comments. Make the argument clearer. Show the evidence earlier. This sounds unlike us. The narrative doesn’t know where it’s going. The example is technically correct but not useful to this audience. The draft explains the product but never establishes why anybody should care.

A good editor can recognise these problems quickly because they have seen versions of them repeatedly. But the correction is usually applied to one document, in one workflow, at one moment.

The writer accepts the change. The document gets published. The comment thread disappears from view.

Then another person makes a similar mistake in another draft and the editor explains the same principle again.

That is not because teams lack guidance. Many have style guides, brand documents and content templates. The problem is that these tend to record the visible rules: preferred terminology, punctuation, formatting, tone of voice.

The more valuable judgement is situational. What makes a strong argument here? What evidence is enough? When does a product claim become too broad? What does a credible executive point of view sound like? Why did this customer example work while a superficially similar one did not?

That knowledge is harder to reduce to a static rulebook. It lives across examples, decisions, exceptions and repeated conversations.

Sandhya’s post treats those conversations as a dataset rather than editorial debris, which is the important move.

The first draft can inherit the previous edits

Editorial processes have traditionally placed judgement late in the workflow.

Someone writes a draft. An editor reviews it. The editor identifies the structural or narrative problems. The draft goes backwards. Eventually it becomes good enough.

Some iteration will always be necessary. Good ideas often emerge through writing, and no system can anticipate every editorial decision.

But a surprising amount of revision is predictable. If an editorial team has corrected the same weakness across twenty drafts, it should not need to wait for draft twenty-one to explain the principle again. The next writer should be able to inherit some of that accumulated judgement before submitting anything.

This is where focused AI agents or skills become more useful than a generic instruction to “improve the writing”.

A useful editorial agent can be given a particular job: test the strength of the argument, check whether the evidence supports the claim, compare the draft against successful examples, identify where product detail has displaced audience value, or challenge whether the opening establishes a real reason to continue.

The goal isn’t automatic approval. It’s a more informed first pass, after which the human editor spends less time correcting familiar weaknesses and more time on the genuinely difficult questions: whether the argument matters, what is missing, what should be cut, what the organisation can credibly say and whether the piece deserves to exist at all.

AI doesn’t remove editorial judgement in this model. It increases the reach of judgement that has already been exercised.

This is close to what I have been building

My own system began from a slightly different problem.

I was producing ideas, frameworks, Field Notes and articles quickly, but I did not want each new piece to begin in an empty chat window. I wanted previous thinking to remain available: the arguments I had already made, the distinctions I cared about, the evidence behind them, the edits that improved them and the contradictions that had not been resolved.

So I started treating GitHub as more than somewhere to store a website.

The repository contains the published work, its relationships, controlled metadata and its revision history. A separate core preserves developing frameworks, research questions, observations and working decisions. AI can work across that context, compare a new idea with what already exists and help decide whether the idea needs a new article, belongs inside an existing one or challenges something I previously believed.

This has changed the practical nature of publishing.

A conversation can surface an observation. The observation can be tested against earlier work. An existing argument can be extended without losing its history. Related pages can be updated. The site can become more coherent as it grows rather than simply becoming larger.

I wrote in Context is capital that AI can make old work more valuable by allowing it to participate in new thinking. Sandhya’s example shows the same principle operating at the level of an editorial team.

The repeated edit becomes a reusable standard.

The strong draft becomes a working example.

The discussion about why something failed becomes guidance available before the next failure.

The organisation does not merely produce better content. It improves the conditions under which future content is produced.

I thought I was improving prompts

The OpenAI example landed because I had already built most of this without quite seeing the whole shape.

In my personal work, Editorial Intelligence OS had shown me that a repository could do more than store finished material. GitHub preserved the arguments, frameworks, changes and relationships. ChatGPT and Claude could work across that accumulated context. A new observation could connect to something written weeks earlier, expose a gap, extend an existing article and leave the system more useful than it was before.

The value was not one spectacular output.

It was that yesterday’s work changed what was possible today.

At Sage, I had been developing a more contained version through customer stories. I initially thought of this as prompt improvement. The first prompt turned a transcript into a draft. It could write competently, but it tended to include every interesting theme. A transcript containing several plausible stories produced a draft containing all of them, without choosing which one mattered most.

So the workflow changed.

An Editorial Brief Generator was added before drafting to propose the story spine, protagonist, strongest evidence, risks and alternative angles. A human editor could accept, reject or reshape that decision before the story was written.

Then an Editorial QA stage was added after drafting. It checked the story against recurring problems involving competing narratives, unsupported claims, quote integrity, generic openings, product-first writing and unresolved threads.

Across fifteen customer stories completed during 2026, those problems stopped looking like isolated edits. They started looking like knowledge about the format.

I documented the failure modes. I added a validation rule: one occurrence remains an observation; a checkable issue appearing across different stories may become a system improvement. A prompt change has to trace back to real work and then prove useful on later work. Otherwise it should be revised or removed.

I thought I was making the prompts better.

What I was actually building was a way for editorial judgement to accumulate.

That is the principle, and the accidental discovery that it was already happening. What it takes to do deliberately is a separate problem, along with the risk that comes with it — a system good at applying yesterday’s judgement can make yesterday’s judgement much harder to question.

Part two picks that up: how the feedback loop closes, why a recurring format is the right place to start, why this is more than a prompt library, and what stops the whole thing hardening into a formula.