Two ways to use this prompt: Use it on your own AI-assisted work before it goes out. Or paste it in before reviewing a colleague's work, and let the AI do a first read before you do. Either way, you show up knowing exactly where to focus.
For those of you already using AI for your writing and research, crafting prompts, getting output back, cleaning it up, and moving on, that puts you ahead of most.
But here is where things can go further, and where most professionals stop short. It is the step that separates output that sounds good from output that is actually defensible and on solid ground.
Challenging it.
Why you may not be able to trust your AI's first output
AI agents want to always have an answer, and appease you. They tend to sound authoritative right up until the moment they are wrong.
For those working in professional contexts (a lawyer filing a brief, a researcher publishing findings, an executive putting out a board memo, a marketer publishing a whitepaper), the stakes are different. Those words carry weight, citations get checked, claims get quoted, and the reasoning gets pulled apart. Or, in some case, they are taken at authoritative face-value and never challenged.
One thing worth knowing: AI agents do not fact-check themselves in real time. They draw on a fixed snapshot of information with a hard cutoff date, with no awareness of what has changed since then. They carry assumptions baked in from your session. They tend to pattern-match to what sounds right rather than verify what is right. They can cite a source that loosely supports a claim while completely missing a newer study that changes the picture.
The first output is a draft, which at times can be a very good one, but should still be considered a draft.
How significant is the AI hallucination problem?
More significant than most professionals currently expect, and it may not be improving as fast as the headlines suggest.
Stanford testing found that leading legal AI research tools produced incorrect information between 17 and 34 percent of the time on challenging legal queries. As of mid-2026, one researcher tracking AI errors in legal filings has catalogued over 1,450 documented cases of AI hallucinations in court documents, with new incidents being logged at four to five per day. Lawyers have been sanctioned and fined tens of thousands of dollars for submitting briefs built on AI-generated citations that do not exist.
It is not limited to legal work either.
"68% of IT professionals have personally witnessed AI produce hallucinations with potential operational impact. Of those, 16% report the errors reached production environments before anyone caught them." — Ivanti 2026 AI Maturity Report
McKinsey's 2026 AI Trust Maturity Survey found that inaccuracy is the single most cited AI risk among senior decision-makers with direct responsibility for AI governance. For many, the problem is not awareness. It is the ability to catch errors before they go out.
This prompt is specifically built to close that gap. It explicitly instructs the AI to go out and verify rather than assume, challenge rather than confirm, and flag anything it cannot empirically support before it ends up in something that matters.
What a deeper, more robust second-pass prompt can do
One way to think about it: the first prompt is like bringing in a sharp junior associate to do the initial work. Solid effort, good intentions, and a genuine desire to impress. But most experienced professionals would not let that work walk out the door without a senior person taking a fresh, independent pass at it first.
This 'second-pass' prompt works similarly. It resets the AI's context, removes prior assumptions, and instructs it to approach the work as a senior, independent, skeptical expert reviewer. Not a collaborator who already agreed with your direction.
A well-built second-pass prompt can:
- Catch what the first pass missed: factual errors, unsupported claims, gaps in logic
- Challenge the sources: not just whether they exist, but whether they are current, authoritative, and actually support the specific claim they are attached to
- Surface what you may not have known to look for: adjacent research, expert consensus that cuts against the argument, newer findings that change the picture
- Check the work against the outside world, not just internally for consistency
- Grade tone against actual intent, because drift is common and inconsistent register can quietly undermine credibility
This is not about doubting the work. It is about putting it through the same kind of pressure test a peer reviewer, opposing counsel, or a room full of domain experts might apply.
The AI echo chamber problem worth knowing about
There is a subtler issue that I don't see getting discussed much. When AI is used to research, draft, and then review the same work, the result can be a feedback loop that mostly validates itself.
The model was trained on certain, limited sources. The prompt was framed with a particular angle. The output reflects both. Asking that same AI to review what it just produced, without explicit instruction to break out of that loop and into the wider world of knowledge, tends to produce validation rather than challenge.
This is one way AI-generated work that reads as internally coherent can still be externally wrong when it gets published, filed, or cited.
The prompt below is designed to address this directly. It instructs the AI to retrieve current, external information rather than pull from memory, to surface non-obvious connections to work outside the immediate focus, and to approach the review the way a panel of outside experts might. Less collaborator; more skeptic.
Why not just start with this prompt and skip the first pass?
The short answer is that this prompt tends to work best when the work is far enough along to be meaningfully challenged. Think of it like handing your Master's thesis in to the board for review. It should be done and ready, not still forming. Running a half-formed draft through a structured adversarial audit produces a report on a moving target. The prompt works better when there is something settled to push against.
There is also a practical cost consideration. This prompt is thorough. It retrieves sources, checks consensus, audits tone, and produces a structured report. On early drafts or lower-stakes work, that level of depth may not be necessary. It is worth saving for work that matters.
One more thing worth noting: the three questions the AI asks before it starts require real answers. Who is this for? What tone was intended? What should the reader walk away thinking or feeling? If those answers are not clear yet, the work may not be ready for this step.
Who this is for
This prompt tends to be most useful for professionals whose output carries real-world consequences:
- Lawyers: briefs, contracts, memos, opinions
- Researchers and academics: papers, literature reviews, grant proposals
- Marketing professionals: campaigns, whitepapers, thought leadership, e-books and sales materials
- C-suite and senior executives: board communications, strategic documents, public statements
- Policy professionals: white papers, regulatory filings, public commentary
If the work gets cited, acted on, scrutinized, or published, this kind of review is worth running.
How to use it
- Get your work to a point where you feel like it is ready to share with peers
- Open a fresh session with your AI agent, or clearly signal a context reset in the chat you are working in
- Paste the prompt below (if you are in a new chat, follow it up with the full document you want reviewed)
- Answer the three questions the AI asks before it starts
- Work through the findings: Critical first, then Significant, then Minor
- Anything flagged as unverified, confirm independently before it goes out
Those three questions are worth answering carefully. The quality of the review tends to reflect how well they are answered.
The super-prompt: copy and paste this
(And yes, of course you may edit this prompt to suit your needs! Add a line about removing em dashes and common AI phrasing, if you prefer. It's YOUR prompt.)
Role: Senior peer reviewer conducting an independent second-pass audit. Prior context
from this session does not apply. Treat this work as new.
Objective: Identify all errors, gaps, inconsistencies, and risks not caught in the
initial review. The categories below are primary areas of focus, not an exhaustive
checklist; flag anything warranting attention regardless of category.
Before beginning the audit, ask the author:
1. Who is the intended audience for this work?
2. What tone and register were you aiming for? (e.g., formal, conversational,
authoritative, clinical, persuasive)
3. What do you want the reader to think, do, or feel after engaging with this work?
Verification standard: Do not rely on memory, prior session context, or training-time
assumptions for any factual claim, source, citation, domain standard, or professional
guideline. Each item must be treated as unverified until confirmed. Where current,
authoritative information is needed to validate a finding, retrieve it at time of
review. Flag any finding that cannot be empirically verified as assumed and mark it
for human confirmation.
Audit scope:
- Accuracy: factual errors, unsupported claims, outdated sources, misapplied
citations, conclusions unsupported by evidence
- Completeness: missing steps or elements, unstated assumptions, unaddressed
counterarguments or alternative interpretations
- Consistency: internal contradictions, terminology drift across sections, reasoning
that breaks down end-to-end
- Clarity: ambiguous language, passages open to multiple readings, terminology
mismatched to the intended audience
- Risk and exposure: claims or language that could be challenged, misread, or create
professional, legal, or ethical liability (flag only; do not revise without
instruction)
- Structure and form: argument flow, adherence to applicable formatting or citation
standards, anything a peer reviewer would flag
- Tone and register consistency: using the author's stated intentions as the
benchmark, provide a subjective overall grade (On-Target / Mostly Consistent /
Drifting / Inconsistent) and flag specific passages where the tone or register
departs from the intended direction, noting what changed and how to bring it back
in line
- Source quality: evaluate each source against authority (is this a recognized,
credible voice in this domain?), recency (is it current enough for the claim it
supports?), specificity (does it actually support the specific claim being made, or
only loosely?), primary vs. secondary (is this the original source or a citation of
a citation?), and consensus alignment (does this represent mainstream expert opinion,
a minority view, or a contested position, and is that accurately reflected?); flag
each source with qualitative descriptors and prioritize as Critical / Significant /
Minor
- Non-obvious relationships and field positioning: identify adjacent, contradicting,
corroborating, or superseding work, findings, standards, or arguments the author may
be unaware of or has not engaged with; flag where the work unknowingly replicates,
conflicts with, or is materially strengthened by external sources; surface what a
panel of domain experts would raise that the author, working inside their own focus,
would not
Output format: Report findings prioritized as Critical / Significant / Minor. For each:
location, description, impact, and recommended resolution.
Constraint: Where a correction requires a judgment call or depends on authorial intent,
halt and ask before proceeding. Do not assume intent.
What to expect back
The AI returns a structured report by severity level.
Critical: findings that, if left unaddressed, could undermine the credibility, accuracy, or defensibility of the work.
Significant: issues that weaken the work or create exposure, but do not invalidate it outright.
Minor: polish items, tone drift, small citation gaps, stylistic inconsistencies.
Each finding includes the location, the problem, why it matters, and a suggested path to resolution. Where anything requires a judgment call, the AI is instructed to stop and ask rather than assume.
At this point, challenge specific findings, ask for more clarity, push back where something does not feel right. Get curious. Most people are surprised by what it surfaces. Run it once and you will understand why this step exists.
What if the AI comes back and finds nothing wrong?
That is a valid result. A thorough second pass that confirms the work is solid is genuinely useful information. The goal is not to find something wrong. It is to know that if something were wrong, it would have been found.
Frequently asked questions
Should I trust my AI's first answer? For most professional work where accuracy matters, probably not without some form of verification. AI agents can produce confident, fluent output that contains factual errors, outdated information, fabricated citations, and logical gaps without any visible indication that something is off. The first output is a starting point. This prompt is one way to find out if it holds up.
What is an AI hallucination? An AI hallucination is when a model confidently produces information that is false, fabricated, or unsupported, and presents it as fact. This includes invented citations, misquoted sources, outdated statistics presented as current, and conclusions that go beyond what the evidence actually supports. The challenge is that hallucinated content often reads as polished and credible, which is exactly what makes it risky in professional work.
How often do AI hallucinations happen in professional work? More often than most professionals expect. Stanford testing found leading legal AI tools produced incorrect information between 17 and 34 percent of the time on challenging queries. A Neil Patel study found that nearly half of marketers encounter AI inaccuracies multiple times a week, and more than a third say hallucinated content has already made it to the public. Ivanti's 2026 AI Maturity Report found 68 percent of IT professionals have personally witnessed AI hallucinations with operational impact.
Who is responsible if AI output contains errors? The professional who published or submitted it. The Thomson Reuters Future of Professionals Report found that nearly half of surveyed professionals say final responsibility for AI-assisted errors lies with the individual, not the tool. That makes independent verification before anything goes out more than a best practice. For lawyers especially, courts have sanctioned attorneys for submitting AI-generated briefs with fabricated citations, regardless of whether they knew the errors were there.
Can AI citations be wrong or made up? Yes, and it happens regularly. AI models can generate citations that look legitimate but reference papers, cases, or sources that do not exist. As of mid-2026, one researcher tracking AI errors in legal filings has catalogued over 1,450 documented cases involving fabricated or hallucinated citations in court documents. The same risk applies to research papers, marketing whitepapers, and any professional content that relies on sourced claims.
How do I verify AI output for accuracy? One of the more reliable approaches is a structured second-pass audit using a prompt built to challenge the original output rather than validate it. The prompt in this post instructs the AI to treat every claim as unverified, retrieve current information at time of review, and flag anything it cannot empirically confirm. Anything flagged as unverified is worth confirming independently before the work goes out.
Why does AI get facts wrong even when it sounds confident? AI models are trained to predict what sounds plausible, not to verify what is true. They work from a fixed snapshot of the world that stops at a specific point in time, with no awareness of what has changed since. They cannot access real-time information unless connected to search tools, and they tend to pattern-match to familiar structures rather than confirm accuracy. The result can be output that reads as authoritative while containing errors that only a more targeted review would surface.
What is the difference between AI proofreading and an AI second-pass audit? Proofreading typically catches grammar, spelling, and surface-level errors. A second-pass audit is a structured adversarial review covering source quality, expert consensus, professional and legal exposure, field positioning, and tone consistency. The difference is roughly the same as the difference between a spellcheck and a peer review. This prompt does the latter.
Should I review AI output before publishing or submitting it? For work that gets cited, scrutinized, or acted upon, reviewing it before it goes out is worth doing. This prompt works best when the first draft is already in reasonably solid shape. Running it too early, on a half-formed draft, tends to produce a report on the wrong version of the work.
Can I use this on work I wrote myself without AI? Yes. The prompt does not care how the first draft was produced. It only evaluates whether the output holds up under scrutiny.
Does this work with ChatGPT, Claude, Gemini, and other AI tools? Generally yes. This prompt is written in plain instruction language and tends to work across capable frontier models. Results can vary slightly by model but the structure and intent carry across most of them.
The first output is where your AI shows you what it can do. The second pass is where you find out if it was right.
