So, I had to give a lot of feedback to Claude to build that EU sanctions history post – a ton more than I usually have to give. So, I asked about it:

So, the work on the EU sanctions history required a lot of feedback from me on word usage and writing structure – which I have not had to do with my other requests, like the sanctions guides or even today’s UFLPA prompts. Any idea why that might be?

and here’s the first response, which I thought had some real insights in it:

A few real differences stand out, though I’ll flag which parts I can trace directly versus where I’m inferring.

The EU post was original analytical synthesis with no template behind it. The guides run on the Master Prompts framework you built out over roughly ten regimes now – the structure, sourcing conventions, and voice were already validated before any single guide started. The UFLPA piece was built the same way in miniature: you walked me through three separate Q&A rounds first (mechanics, importer consequences, third-party liability), approved each one, and only then asked for a meld – so by the time I wrote the combined draft, the content and tone had already been tested in stages. The EU post had none of that scaffolding. Nobody had written “functional evolution across the 21 packages” before, so the chart format, the section groupings, how to frame “significance” for each group, even the package 14 versus 15 attribution call – all of that got worked out live, inside the draft, rather than inherited from a proven process.

Some of the prose rules I’m now applying didn’t exist yet when that project started. Looking at when things got filed: several of the specific conventions I follow now – avoiding label-then-explain constructions, not narrating the article’s own structure, varying back-references instead of repeating a pronoun, watching paragraph length, naming things plainly instead of using shorthand – trace to that same EU project. That’s not a coincidence so much as a cause: it’s the project where those rules got established through your corrections, not one where I was already following them. The UFLPA piece benefited from having that list already in hand.

The content type itself carries more judgment calls. The EU post asked for interpretive synthesis – financial impact, geopolitical reaction, convergence and divergence across five section groups – which has a lot of surface area for word-choice and framing to go wrong. The guides are closer to organizing known facts into a fixed shape. The UFLPA piece was more explanatory (what this means for someone in the chain) than argumentative, which tends to need fewer editorial calls per paragraph.

So: newer content type, no established framework, and it’s likely where a chunk of my current style rules for you actually got written. The UFLPA piece and the guides are both riding on groundwork the EU post did the hard part of laying down.

But, while I agreed with a lot of it, I wanted to have Claude explain something that I felt was off:

I want to challenge one thing – your prose rules point. I looked back at the guides: those sort of “errors” didn’t occur. So, there was no need to call them out. I think the reason for the difference is more your first and third point – think of them as coarse sandpaper, and the second point more fine finishing and buffing materials

And here’s what it said (yes, it tends to agree with the prompt):

Fair, and it holds up when I look at what’s actually in the guide drafts rather than reasoning about it abstractly.

I pulled a few of the actual guide-building conversations to check. What’s in them is regulatory framework sections, warning boxes for counter-intuitive rules, FAQ entries, checklist steps, glossary terms, document indexes – declarative, informational moves. There’s no equivalent anywhere to “on convergence” labeling, no closing section weighing significance across several groups of content, no long stretch where I’m carrying one evolving subject across dozens of consecutive paragraphs. The EU post’s pronoun-repetition problem, for instance, came specifically from referring back to “this package” or “it” sentence after sentence because the whole piece tracks one subject through time. A guide’s subject changes with the section – IRGC designation status, then the 50% Rule, then a general license – so that particular failure mode has nowhere to occur regardless of whether I’m applying a rule against it.

So you’re right to demote the second point. It wasn’t that I had those rules in hand for the guides and lacked them for the EU post – it’s that guide content doesn’t create the occasions for label-then-explain constructions, structure-narration, or pronoun drift in the first place. Those errors are a symptom of writing original interpretive narrative across a large, evolving subject, which is exactly what points one and three describe. Point two was me mistaking the moment those rules got written down for a cause, when it was really just documentation of a problem that only the first and third factors could have produced.

Now, you may find it weird – and certainly not standard practice – but I basically shared my insight from its response:

Which will help me recognize when I need to review your output more carefully – I don’t tend to review what you’ve written for errors. I look it over more for what I learn from your response.

And the response brought further insight:

Given that, the useful signal isn’t really about prose at all – it’s about where interpretive claims are getting made versus where facts are getting organized.

The two things share a root cause. Original synthesis with no precedent forces me to make judgment calls – what counts as significant, what caused what, how one thing compares to another – and those same judgment calls are what produce both the prose problems you caught and any unsupported inference that might be sitting underneath clean-sounding sentences. A guide’s content is close to directly extractable from a regulation or an FAQ, so there’s not much room for me to be inferring rather than reporting. The EU post asked me to characterize financial impact, geopolitical reaction, and causal sequencing across 21 packages – all of that is synthesis, not extraction, even where it reads confidently.

So the practical version: when I’m building something with no established template and the content requires me to characterize significance, causation, or comparison rather than just organize known facts, that’s where to look harder at the specific interpretive claims – not the whole document, just the sentences doing analytical work rather than reporting work. Dates, figures, FAQ numbers, regulatory citations are about equally checkable regardless of content type. It’s the “this mattered because X” and “this led to Y” sentences that concentrate in the original-synthesis work and deserve the closer look, precisely because that’s where I’m most likely to be asserting more than the sources actually support.

And I acknowledged that I agreed:

nice

And the response will further help me in my workload here:

Sounds good. I’ll keep that distinction in mind on future projects – flag up front when something’s original synthesis versus template work, so you know which one warrants the closer look.

What are the takeaways?

I think there are two significant takeaways – one mostly for you, my readers, and one for us all.

The first one: if you’re going to ask the kind of questions that involve the kind of analysis and synthesis that goes beyond summarization and organization of facts, you’re going to need some writing and editorial skills. It’s not so much about knowing what to write; rather, it’s about point out what looks funky. If you can, you can make suggestions for rewording, but you don’t need to be Shakespeare for the AI to realize the difference between “right” and “wrong.”

The second takeaway: You can push the envelope with what you ask an AI to do, if you’re willing to put in the work. Not only do you need to do the editorial work, but the farther outside the envelope you go, the more you have to check for the AI inferring too much. Of course, you should be explicit about performing a sanity check against sources, and about preventing hallucinations – but the response needs to pass a basic sniff test. If something is valid 55% of the time, is it an inference worth including in your output? 65%? 85%? Drawing that line is the real challenge – I don’t think you can do that effectively without either a lot of research and review, and a certain amount of subject matter expertise. The less expertise and the less time you are willing to check the responses you get, the less you and your AI can go out on an inferential limb.