My editing of Claude’s responses is pretty minimal, and I want to explain it, in the service of full transparency.

What I do edit out is basically three things:

• In a chatbot, any process-type discussion (couldn’t access this document, now fetching this) appears as part of the response block – and gets picked up if you hit the Copy button. The process of how a chatbot tries to gather information, as a general matter, doesn’t add that much value, IMHO, to most of my intended audience – even the more AI-curious. So, I leave these out. Ditto for any discussion as it refines its thinking – what we largely care about is the final response. A good example is the (upcoming) Burning Questions post on the “secondary risk” wording in OFAC listings, where Claude prefixes the final response with this (none of which appears in my post):
  • This is the sharper version of the “Burning Questions” angle from before — whether the field is being read as a rating rather than a citation, and whether that misreading clusters anywhere. Let me search for evidence rather than reason from priors, because the honest answer may be that this specific behavior is under-documented.
  • That gave me adjacent material (screening mechanics, one vendor that tags secondary sanctions as a labeling feature) but not direct evidence of firms misreading the field as a rating. Let me look specifically at the de-risking / over-compliance pattern, which is where any geographic or asset-class clustering would show up.
  • I now have solid material on the de-risking pattern and where it clusters, but I want to be honest about a distinction: the question asks specifically whether firms treat the wording as a rating, and most of what I’ve found is about firms reacting to the underlying risk. Those aren’t the same thing, and I don’t want to conflate them. Let me do one more targeted search on the data-field-as-signal angle.
  • I have enough to give an honest, well-bounded answer.
• Sometimes, as part of its source checking, Claude will point out things where the difference between its initial drafting and its checking – semantically and in a practical sense – is not worth correcting, or where the discussion of the review itself doesn’t illuminate the response. Those I remove. If there is a real difference, by the way, I have Claude correct it – so I don’t post the original response.
• Often, Claude will propose I ask a follow-up question. These I remove.

Now, I do leave in some of Claude’s commentary from time to time. I do that for a number of reasons:

• I think the reasoning, and the natural language production of the “conversation”, is fascinating
• I think the back and forth of postulating something, posting a caveat and then proceeding, is more nuanced and informative to the audience (and, to be perfectly frank, to me as well)
• As much as this is a sanctions blog, it’s also a “AI in sanctions” blog. Some of these peeks behind the curtain inform us as we make decisions about how to use AI in our field

As always, I invite your comments, criticisms and suggestions.