Here's the mistake almost everyone makes with an earnings call. They paste the transcript into Claude and type "summarize this." Then they get back ten bullet points that read like the press release. Revenue up. Margins stable. Management confident. None of that is research. That's a rewrite of what the company already told you. And the failure mode is brutal, because the thing that actually moves your model is rarely in the summary. It's in what changed.
So let me show you the workflow I use instead. The goal is simple. You're not summarizing the call. You're building a change log. You want changes, contradictions, and risk language. Nothing else.
Step one. Get two transcripts, not one. The current quarter and the prior quarter. Both calls, prepared remarks and Q&A. You can pull these from the company's investor relations page or your transcript provider. Drop both into Claude in the same conversation. This is the part people skip, and it's the whole point. You cannot find a change with one quarter. Change is a comparison.
Step two. Do not ask for a summary. Give it a job. My prompt is roughly this. "Here are two earnings call transcripts for the same company, this quarter and last quarter. Compare them. Build me a change log with three sections. One, language that changed on the same topic between the two calls. Two, places where management contradicts something they said last quarter or earlier in this same call. Three, every instance of hedging or risk language, things like 'we're monitoring,' 'softer than expected,' 'pushed to the right,' 'more cautious.' For each item, quote the exact sentence from each call and cite which quarter and which section it came from."
That's the unlock. You're forcing Claude to anchor every observation to a quote and a source. No paraphrasing. No vibes. If it can't quote it, it doesn't go in the log.
Step three. Read the contradictions section first. This is where the signal hides. Last quarter the CFO said pricing was holding. This quarter the word "promotional" shows up four times. Last quarter guidance was "on track." This quarter it's "we now expect." Those shifts are small in word count and large in meaning. Claude is genuinely good at catching them because it's reading both calls at once and it doesn't get tired in the Q&A section the way you do at 5 p.m.
Step four. Work the risk language. Pull every hedge into one place. Then ask a follow-up. "Of these risk phrases, which ones are new this quarter versus repeated from last quarter?" New hedging is the tell. A phrase management used last quarter and dropped, or one they introduced this quarter, tells you where their own confidence moved. That's a question for your model and your IC memo. It is not a conclusion. It's a place to dig.
Now the part that matters more than the prompt. The controls. Claude will occasionally attribute a quote to the wrong quarter or smooth two sentences into one. So every line in that change log gets verified against the transcript before it touches your memo. That's the rule. Four things on every item. The source. The date or quarter. The assumption you're making from it. And a human review. You. Not the model. The model structures the reading. It does not reach the judgment.
A few things I've learned running this. Keep the transcripts clean. Strip the operator boilerplate and the safe-harbor statement at the top, because that safe-harbor language is full of risk words that aren't signal, and it'll pollute your hedging section. Run prepared remarks and Q&A separately if the call is long, because the interesting contradictions usually live in Q&A where management is off-script. And always quote, never trust a paraphrase. If you're going to put something in front of your PM, you should be able to point at the exact sentence.
Why does this work better than a summary? Because a summary compresses toward the average, and the average is what the company wanted you to hear. A change log does the opposite. It surfaces the deltas, the places where this quarter doesn't match last quarter. That's where your edge is. Not in knowing revenue grew. Everyone knows revenue grew. In noticing that the language around demand got two clicks more careful while the headline number stayed fine.
You can run this whole thing in about fifteen minutes once you have both transcripts open. Paste two quarters, ask for the three-section change log with quotes and sources, read contradictions first, isolate the new risk language, then verify every line by hand. That's the loop.
One more time on the boundary, because it's the most important part. This is a research process, not a recommendation engine. Claude is reading transcripts faster than you can and organizing what it finds. It is not telling you what the stock is worth and it is not telling you what to do. You build the change log, you verify the quotes, you form the judgment, you write the memo. The tool handles the reading. You handle the decision.
So next earnings season, before you type "summarize," try the other prompt. Two quarters, one change log, quote everything, verify everything. Find the changes, the contradictions, and the risk language. That's the call within the call. This is educational only. It's not investment advice. Markets involve risk, verify your data, and do your own work.