- Ask for the batch, not the lookup. The research platform already had a tool that checks a whole list at once. I had to ask for it.
- The citator checks the case you hand it. My AI pointed it at the wrong decision once and at the wrong dates twice. Checking each case before trusting the answer caught both.
- The flag says look; the opinion says what. A “criticized by” label turned out, on a full read, to be a certified conflict headed to a state supreme court.
I don’t run a law practice. I build the AI operating systems that run other people’s — AI-enablement for legal, full stop. This month I am trying Lexis+ with Protégé, and the work I gave it is the last step before the 50-state brief publishes: confirm that every court decision the brief cites is still good law. I did not do that step by hand. Claude Code, the AI assistant I build with, worked in my signed-in browser session one page at a time, built the lists, ran the searches, read the reports and kept the record, while I made the calls. It changed how the whole job runs. It also made three mistakes that would have reached a public page, and the reason they did not is the useful part. Notes twenty-five through twenty-seven.
25. Ask for the batch, not the lookup
The brief cites 84 decisions across 47 jurisdictions. The last edition’s check ran one citation at a time on a different citator: type the citation, read the signal, open the report, write it down, next. Seventy-one citations took several sittings. When I switched to Shepard’s this time, my assistant proposed the same loop, at a minute or two a citation. I said no, and asked it to find a way to run them all at once. It already knew one: Shepard’s BriefCheck takes a list of up to six hundred citations and returns a report on each. The first run took 51 citations and came back as a single 337-page report. The assistant found the citations for 39 more with ten searches, each pairing a case name with its decision date, and a second run returned another 207 pages. Seven stragglers it checked on each decision’s own page. All 84 were done the same evening: 83 with no negative treatment, one flagged.
The lesson is not about one product. Almost every research platform you pay for has a batch mode for the step you are doing one at a time, built for exactly this: a brief, a memo, a set of authorities that all need the same check. The default an assistant reaches for is the careful, slow loop, because that is what it can see you doing. Ask for the batch. One thing does not change: a batch report’s first screen is a column of signals, and a signal is not a finding. The value of getting the full reports in one delivery is that every one of them can actually be read.
26. The citator checks the case you hand it
Twice that evening, the problem was not the citator. It was what my assistant handed it. Kansas came back with a red warning: negative treatment. The brief cites a 2026 Kansas Supreme Court decision, and its citation line reads “reversing” the Court of Appeals decision below. The assistant’s citation extractor had pulled the reversed lower-court citation out of that line, so the warning was correct about the wrong case: it said the Court of Appeals decision had been reversed, which is exactly what the brief says. Then, building the searches for the second run, it took the date a trust was signed, from the case name, as the date two cases were decided.
Neither reached a page, because each was checked against the case the brief actually cites: the search dates before the searches ran, and every report’s caption, court and decision date before any result was recorded. That step is dull and it is the one that matters. A citator answers the question it was asked. A clean signal on the wrong case is worse than no check, because it looks like a check. So confirm the identity before you read the signal: same parties, same court, same date. If your tools extract citations for you, assume the extractor will sometimes grab the case being reversed, affirmed or distinguished instead of the one you meant.
27. The flag says look; the opinion says what
The one flag was a Florida will-contest decision from the Fourth District, marked with a caution: criticized by a Second District zoning decision from late September. The brief already carried a caution about it, written during research, calling the later decision unreported and saying it criticized the standing test the will-contest decision used. I had my assistant pull the later decision from the court’s own site and read all forty-five pages. It is a written opinion. The Second District expressly refuses to adopt the federal standing standard that several district courts, the Fourth among them in the will-contest decision, have drawn from a footnote in a 2004 Florida Supreme Court decision; it certifies direct conflict with the will-contest decision on that point; and it certifies to the Florida Supreme Court the question whether that footnote replaced Florida’s own standing standard.
“Criticized” and “certified conflict with a question pending above” are not the same warning. For a lawyer about to argue standing in a will contest, the second means the state supreme court may soon decide the very test the case applies. The signal’s label is the citator’s summary of a relationship; only the citing opinion says what the relationship is. The brief kept the will-contest decision, with the caution rewritten to say exactly that. What remains unsolved: five of the 84 decisions were only days to a few months old, and nothing had cited them yet. For those, “no negative treatment” means only that no court has had time to say anything at all.
These notes aren’t a practice diary — they’re a record of what I keep finding while building AI systems for legal work, written for the solo and small-firm attorneys those systems are for. This one you can act on tomorrow: put your next brief’s citation list through your citator’s batch tool, then open the citing opinion behind every flag it raises.
Field Notes are written for licensed attorneys and are not legal advice. Mike Moss is a Utah-admitted attorney doing AI-enablement work — not operating a law practice.