The Bluff Filter is the paste-in version of this. It’s free →
// On this page
Most AI earnings content stops at extraction. The numbers, the headlines, the ratios. Every general-purpose model handles that well enough, and if you’re deciding which one to use, the comparison piece on the best AI tools for earnings analysis sorts that out. These five AI prompts for earnings call analysis cover the harder half, which is language rather than data. They all run after the release. If you’re after the one that runs the morning before, so you’ve pre-committed your sell, add and hold triggers before management speaks, that’s the pre-trade prompt.
The harder half is reading the language. Not what management said. That’s what every transcript service already summarises. What they committed to, what they hedged, what they conspicuously didn’t address, and how their framing has drifted since last quarter. It’s the language layer of an AI quality of earnings review: accrual quality and the free-cash-flow gap are the accounting side, and what management didn’t quantify is the disclosure side. That’s what these prompts get at.
I held META into the Q1 2026 release. Revenue came in at $56.3bn, up 33% year on year, and the headline was strong. Capital spending guidance, the money going into data centres and chips, was $125–145bn for the full year, given without a payback timeline. The stock fell roughly 8% in the session that followed.
The numbers weren't the problem. The omission was.
Every prompt below assumes you've already pasted the management commentary into the session. The model can't see the live filing, and any prompt that lets it pretend otherwise hands you confident fiction. That's what I got the first time I asked Perplexity to "analyse the last META call" without pasting anything.
All five follow the Prompt Stack: SCOPE, FILTER, RISK, VERDICT. What each of those four is doing sits at the foot of the post, if you haven’t met them before.
Zuckerberg used the word “return” without producing one
Before you’ve read a word of what management said, list what a careful investor would expect to hear addressed. Then check whether each item was covered, hedged, or absent. Management can’t hide what they don’t mention, but a casual reader notices presences, not absences.
I ran this on the META Q1 2026 commentary. The model flagged quantified AI ROI as the most conspicuous absence: Zuckerberg used the word “return” without producing one, while adding $10bn to an already-record capex line.
SCOPE: You have not seen the transcript yet. For the prediction, work from what you can actually source about [COMPANY NAME] and the sector and label it as a prediction, not fact; then check it against the commentary I paste below, working only from that text. Predict what should be in the results, then check what is.
FILTER: Based on what you know about [COMPANY NAME]‘s business, the prior quarter’s commentary, and the sector context, list four to six topics or metrics that a careful investor would most expect management to address this quarter. Be specific: “margin trajectory in segment X” not “performance”.
I will then paste the management commentary below. For each of your four to six topics, mark it as ADDRESSED (with a specific number or commitment), MENTIONED (raised but without specificity), or ABSENT (not raised at all).
[PASTE COMMENTARY HERE]
RISK: For any absent topic, suggest one reason management might have chosen not to address it. Distinguish “we don’t have news” from “we have news we’d rather not lead with”.
VERDICT: Name the single most conspicuous absence and one sentence on whether it changes how the headline numbers should read.
Apple committed to exactly two things about its future
Confident-sounding language splits in two. Genuine confidence has at least one specific number, timeline, or falsifiable commitment. Performative confidence has none. Both sound the same in a transcript, and only one of them’s a commitment. This is one job AI does well, because there’s a clean linguistic signature on each side.
I ran it on Apple’s most recent services commentary. List A came in at 13 statements, List B at 11. Strip out the backward-looking numbers any CFO has to report and what’s left about the future is dominated by List B.
Management committed to exactly two specific forward-looking things. Everything else about the future was mood music.
Services growth next quarter, at a similar rate to the March quarter once you strip out the tailwind from currency moves.
Apple Maps ads, launching this summer in the US and Canada.
Mood music. It reads like a plan, but there's no specific number, no timeline and nothing falsifiable in it.
SCOPE: Work only from the management commentary I paste below; don’t bring in outside facts from memory, and say so if the text doesn’t show something rather than fill it in. You are assessing whether the language reflects genuine confidence or performative confidence, not whether the business is doing well.
FILTER: Below is the management commentary section from [COMPANY NAME]‘s [QUARTER] results. Produce two lists.
LIST A - GENUINE CONFIDENCE: statements with at least one of a specific number, a specific timeline, a falsifiable commitment, or a comparison to a prior stated target. Example structure: “We delivered X% growth in segment Y, ahead of the Z% guidance given in [prior period].”
LIST B - PERFORMATIVE CONFIDENCE: statements that sound confident but contain no specific commitment, no timeline, and no falsifiable claim. Example structure: “We remain well-positioned to capitalise on the opportunities ahead.”
[PASTE COMMENTARY HERE]
Count both lists. If List B is materially longer than List A, name the topics where confidence is performative. Those are the topics management is uncertain about and is choosing not to say so.
VERDICT: One sentence. Does this commentary commit management to anything specific, or does it preserve their optionality?
Consensus said about $122.6bn. Guidance topped out at $145bn.
The gap between management's language and consensus's mid-point is where stock reactions live.
A reader who doesn’t track consensus loses the magnitude of any surprise. They hear “increased investment level” and they don’t hear that the top end of a guide sits well above what analysts modelled. Give the AI the consensus number and the language reads differently.
I was holding META into the Q1 2026 release. FactSet consensus going in was around $122.6bn of full-year capex. The $125–145bn band came in above that, and the high end was the part the market punished. The language framed the raise as “investment in AI capability” without a payback horizon.
The gap to consensus was the thing the prompt could have caught before the stock moved. It doesn't need a forecast or a model. It needs one number typed in above the commentary.
SCOPE: Work only from the commentary I paste below and the consensus figures I give you; don’t fill in numbers from memory, and if something you’d need isn’t here, say so rather than guess. You are checking whether your model assumptions need to change after this commentary.
FILTER: Before reading the commentary below, here is the consensus picture for [COMPANY NAME] going into this release:
- Consensus revenue [PERIOD]: [VALUE]
- Consensus margin: [VALUE]%
- Consensus capex / opex / [KEY LINE]: [VALUE]
- Consensus guidance midpoint: [VALUE] Source: [BLOOMBERG / VISIBLE ALPHA / IBES / etc.] as of [DATE].
[PASTE MANAGEMENT COMMENTARY HERE]
For each consensus assumption, identify any phrase in the commentary that supports it, contradicts it, or implies a different number. Quote the phrase. Quantify the implied gap where you can.
RISK: Identify the single consensus assumption most likely to need a revision after this commentary, and the direction.
VERDICT: One sentence. Does the commentary read as ahead of, in line with, or behind consensus on the line item that matters most for the stock?
Same ETH programme, two different sentences
This is the move only AI can do at speed.
A human reader can hold one prior transcript in their head. The model can hold four.
Same activity, different framing usually signals something: confidence rising or falling, with the language as the early tell. BMNR’s framing around its ETH buying programme shifted between Q4 2025 and Q1 2026.
Read as aggressive accumulation. That's the earlier call, on the same buying programme.
Reads closer to disciplined positioning. Around the time the buying pace was publicly described as moderating toward the 5% supply target.
Same programme, different sentence. I’ve been reading every BMNR release for this kind of shift since the shares moved up to the main NYSE exchange.
SCOPE: Work only from the two quarters of commentary I paste below; don’t bring in outside facts from memory, and if the text doesn’t show something, say so rather than fill it in. You will compare how [COMPANY NAME] management framed a specific topic across two consecutive quarters.
FILTER: Topic to compare: [TOPIC - e.g. “margin trajectory in segment X” / “AI capex strategy” / “subscriber growth”].
[PRIOR QUARTER LABEL - e.g. “Q4 2025 commentary”]: [PASTE PRIOR QUARTER LANGUAGE ON THE TOPIC]
[CURRENT QUARTER LABEL - e.g. “Q1 2026 commentary”]: [PASTE CURRENT QUARTER LANGUAGE ON THE TOPIC]
Produce a side-by-side analysis covering:
- Specificity: did the language become more or less specific?
- Time horizon: did the timeline shorten, lengthen, or vanish?
- Commitment level: hedge words added or removed?
- Frame: did the topic shift from achievement (past) to intention (future), or the other way?
RISK: Name the single most material shift and one observable signal in the next quarter that would confirm whether the shift reflects a real change in trajectory or just a different draft of the same script.
VERDICT: Rising confidence, falling confidence, or essentially unchanged framing? One sentence.
The one number Apple won’t put on the call
The hardest question management didn’t have to answer, usually because the analysts weren’t aggressive enough, or because the question got politely deflected. That’s the one worth holding in your head before you decide whether to act.
I ran this on Apple’s most recent services commentary. The model picked the Google search licensing payment, reportedly around $20bn a year and flowing largely to profit, as the structural question.
Predicted response from the CFO: total non-disclosure on Google economics, which is itself the tell.
SCOPE: Work only from the commentary and any Q&A I paste below; don’t bring in outside facts from memory, and if something you’d need isn’t here, say so rather than guess. You are the most sceptical short-side voice covering [COMPANY NAME]: you believe the bull case is overstated and you have one question on this call. Make it the right one.
FILTER: Below is the management commentary and any Q&A excerpts you have. Read everything once before formulating your question.
[PASTE COMMENTARY AND Q&A]
Your question must:
- Refer to a specific claim or number management made on this call, not a generic concern
- Identify the assumption the claim depends on, not the claim itself
- Be a question a fund manager would ask on the next call, not theatre, not gotcha
RISK: Predict how management would most likely respond to your question, in their style. Then identify the part of their response that would itself be evasive.
VERDICT: State your question in one paragraph. Then state, in one sentence, what their answer would need to contain for you to leave the holding alone, and what would make you sell some.
Where these prompts fall short
- Pasted text is the limit. The model can't assess anything you don't hand it, and if the commentary runs long and you copy the highlights, you've already done the sceptical sort yourself. I've caught myself doing exactly that.
- Tone scoring is unreliable. "We are taking a conservative approach" reads as bad news to a generic AI model, and it's often a sign of sensible spending in context.
- None of them predict anything. No prompt here calls the next earnings release or the share-price reaction. They read what's in front of you.
- Rehearsed restraint looks like real reticence. A CEO who's naturally careful with language will look hedged on Prompt 2 even when the business is performing well.
The fix on all four is the same. Run more than one prompt and weigh the outputs against the numbers. Every pass works better when the model has been told to separate facts from inference, which is the single prompt change that made AI analysis worth using and the foundation under all five of these.
What I run on results morning
- Numbers first. Perplexity does the fastest live retrieval, and Claude reads the deck more carefully. The tools comparison has the per-task breakdown.
- Prompt 1, before I've read what management said, while my expectations are still my own.
- Prompts 2 and 3 on the prepared remarks, once I've got the commentary open.
- Prompt 4 if I've got the prior transcript handy. For META, BMNR and Apple I always do.
- Prompt 5 last, before deciding whether to act, which feeds into the five questions I ask AI before any trade.
The three red-flag patterns that recur across calls are the companion check: one prompt, cross-call language tells rather than per-call analysis, run on the same pasted remarks. Most of this runs without paying anyone, and the free-tools audit maps what each tool gives you before the paywall appears.
The whole sequence takes about 20 minutes if you’ve got the prior transcript open in another tab. Twenty minutes isn’t fast, and speed was never the point. The numbers tell you what happened and the language tells you what they think happens next.
Why every prompt carries the same four labels
Skippable if you’ve already met the Prompt Stack. All five above follow it, for the same reasons every other post on this site does.
- SCOPE fences the model to the text you paste, and lets it flag what it can't verify.
- FILTER forces real text in, so it isn't working from memory.
- RISK is the bit retail readers skip.
- VERDICT stops it handing back an "it depends" non-answer.
The short version
What worked: Five prompts built around specific tells: omissions, performative confidence, gap to consensus, language shift, the unasked question. Each one’s copy-paste ready, and each runs against pasted text rather than invented numbers.
What didn’t: No prompt here predicts the share-price reaction. None of them replaces reading the document. Tone-classification on financial language is unreliable, and naturally restrained CEOs will look hedged on Prompt 2 even when they shouldn’t.
Bottom line: These are the prompts I run on real earnings releases I hold positions into, and they’ve caught omissions and shifts I’d have missed reading the transcript once.
Ben tests how far you can trust the main AI assistants, and publishes exactly where they get things wrong. Every post here is a first-hand test with the receipts, including the times a tool simply wasn’t worth the trust. About Ben →
The site tests how far you can trust the main AI assistants, on real decisions. Start with the Prompt Stack for the four-stage framework, free and ungated, or the Bluff Filter for the paste-ready version with a real before and after.