The easiest way to misuse AI is to let the same workflow generate the work and then treat its own self-review as validation.
That sounds obvious when stated plainly. Humans do not usually treat self-review as independent validation. We do not ask the author of a memo to be the only reviewer, the only fact-checker, the only critic, and the final judge of whether the memo is ready for external use. Yet that is often exactly what happens in AI-assisted work. A model drafts the report, performs a quick self-check, fixes whatever it notices, and then presents the output as if the review step has meaningfully reduced the risk.
It may have reduced some risk. It may catch obvious inconsistencies, missing sections, formatting problems, or simple internal contradictions. But it is not independent validation.
The creator prompts we have looked at so far can produce useful work. They can even include a non-independent inline self-check. But a self-check is not a Fresh-Eyes audit. It is still part of the creation process. The model is reviewing work produced inside the same context, under the same assumptions, often with the same blind spots that shaped the draft in the first place.
The Fresh-Eyes Validator prompt in my SIYOM prompt pack exists to create a separate audit layer. Independent here does not mean formal third-party assurance, legal review, compliance review, or guaranteed source verification. It means role-separated review: a separate prompt, run against the finished artifact, with instructions not to create, repair, defend, or rewrite the work.
Its job is narrower and more important: audit the finished work as written and tell the user whether it is structurally sound, epistemically disciplined, logically coherent, sufficiently supported by the evidence presented, and fit for its intended use.
Not every AI-generated note, outline, or internal draft needs this level of review. The point is that consequential artifacts should not move from generation to reliance merely because they read well. The point is to decide whether the artifact is sound enough, clear enough, and sufficiently supported to keep moving, or whether it needs repair before you rely on it for further work.
Self-Check Is Not Validation
The Company Report prompt and the Dossier prompt both include inline self-checks. Creator prompts should be asked to check their own work before producing a final artifact. They should look for missing sections, internal inconsistency, unsupported claims, identity confusion, and obvious failure to follow the output contract.
But that does not make the self-check independent. Self-checks happen inside the same generation frame. They share the same prompt context. They may rely on the same mistaken assumptions. They may miss the same ambiguities. They may polish flawed conclusions instead of challenging them. They may turn a weak inference into a cleaner sentence, which is not the same thing as making the inference stronger.
This is why the Fresh-Eyes Validator is a separate prompt, not just a section inside the creator prompt. It treats the finished artifact as something produced by another agent. It does not assume the upstream work was correct. It audits the artifact itself first, then uses any provided upstream context only to test consistency, identity integrity, and contextual fit.
Validation Is Not Revision
A validator should not be a rewriter. This is one of the most important constraints in the Fresh-Eyes prompt. The validator does not repair the document. It does not provide replacement prose. It does not generate a revised artifact. It does not quietly turn critique into edits. It diagnoses what is wrong, what matters, how severe the issue is, and what kind of corrective work is needed.
That may feel inefficient if you are used to asking AI to “review and improve” something in one pass. But that apparently efficient instruction is exactly how validation, critique, and revision get collapsed into one opaque pass. When review and revision collapse into a single step, the model can make changes the human never explicitly approved. It can “fix” a problem by introducing a new one. It can smooth over uncertainty instead of surfacing it. It can translate a concern into prose before the user has decided whether the concern should change the artifact at all. Fresh-Eyes validation should make the user smarter about the artifact before anyone starts rewriting it or critiquing it.
That is why the output is a validation audit, not a revised document. It separates blocking issues, detailed findings, prioritized corrective guidance, a compliance scorecard, and validator sign-off. The user can then decide what to do next: accept the artifact, repair it, run a Red Team critique, gather more evidence, or abandon the output.
Audit Only What Is Present
The Fresh-Eyes Validator is also intentionally audit-only. It does not conduct new research. It does not introduce new sources, new citations, new facts, or new analysis beyond what is necessary to evaluate the artifact as written.
That constraint can feel counterintuitive. If a source is missing, why not go find it? If a factual claim looks suspicious, why not verify it? If a section is weak, why not fill the gap? Because that is not this prompt’s job.
The validator’s job is to tell you whether the artifact has enough support on its own terms. Where evidence is missing, it should flag the absence. Where citation discipline is weak, it should say so. Where a claim is too precise for the support shown in the document, it should identify the risk. Where context is missing, it should note that validation confidence is reduced.
Fresh-Eyes is a source-discipline screen, not full source verification. Because it does not conduct new research, it can flag missing citations, weak support as represented in the document, suspicious precision, and claim-source mismatch risks. It cannot certify that every external source actually says what the artifact claims unless those source contents are provided.
That makes the audit cleaner. It also prevents the validator from laundering an under-supported artifact into apparent adequacy by doing new work that the original artifact failed to do.
When Fresh-Eyes flags a claim as unsupported, outdated, suspiciously precise, or dependent on missing context, the next step is not for the validator to patch the artifact. The next step is human review, source retrieval, external verification, or a revised creator prompt with better inputs.
The Fresh-Eyes question is not: “Can we rescue this?” The question is: “Is this artifact sound as written?”
What Fresh-Eyes Looks For
Fresh-Eyes is less about finding typos than finding category errors: wrong identity, wrong frame, wrong evidence standard, wrong confidence level, wrong audience, or wrong decision context.
It looks for identity and disambiguation problems. Is the subject of the document clearly identified? Has the artifact blended multiple people, companies, subsidiaries, regions, or related entities? Does any provided upstream identity context match the document itself?
It looks for entity classification and framework fit. This matters especially for company and organization reports. A public company, private company, and non-profit organization should not be analyzed with the same objective function, disclosure assumptions, or evidence standards. If the artifact uses the wrong framework, the output may be polished but analytically misaligned.
It looks for structural and scope compliance. Does the artifact include the major sections required for its type? Does it follow a coherent structure? Does it maintain appropriate scope, or has it wandered into unsupported digression?
It looks for citation and source integrity. Are non-trivial claims cited? Do the claims appear aligned with the support represented in the document? Are there suspiciously precise statements without clear evidence? Are weak sources being asked to carry strong claims?
It looks for epistemic discipline. Does the artifact distinguish verified fact, labeled inference, and unknown? Are inferences anchored to evidence? Are confidence labels used where they should be, and are they calibrated rather than decorative? Does the artifact avoid fabricated precision and narrative leaps?
It looks for hallucination and fabrication risk. This does not mean the validator can prove every issue without external research. It means the validator should identify warning signs: invented-sounding facts, inconsistent claims, implausible source use, unsupported synthesis, or confidence that outruns the evidence presented.
It also looks for logic, coherence, and audience fit. Do the conclusions follow from the evidence? Are recommendations supported by findings? Does the tone and emphasis fit the stated purpose and audience? Is the artifact useful for the decision it is supposed to support? Resolving questions like these are the difference between an artifact that sounds good and an artifact that deserves to keep moving.
Examples of What It Can Catch
A Fresh-Eyes audit can catch simple problems, but its real value is in catching problems that are easy to miss when the draft sounds fluent. It may catch a dossier that uses the right person’s LinkedIn profile but mixes in publications from someone else with a similar name. It may catch a company report that treats a private company as if it had public-company disclosure obligations. It may catch a polished strategy implication that is really a single-signal inference dressed up as a conclusion. It may catch a recommendation that is plausible in general but not actually tied to the stated Requesting Party or intended use.
It may also catch softer but still material failures. The artifact may be structurally complete but strategically unhelpful. It may cite sources but fail to use them in a way that supports the claim. It may separate fact and inference in one section, then quietly blur them in the executive summary. It may include a long list of risks without explaining which ones matter for the decision.
Fresh-Eyes is not infallible. A validator prompt can miss the central issue, especially when the missing evidence or context is not present in the artifact. That is why the output is an audit aid, not a final authority. These are exactly the kinds of issues AI-generated work can hide behind competent prose. Fresh-Eyes exists because polished language is not proof of sound analysis.
How I Use It
I use the Fresh-Eyes Validator when I have a finished artifact that sounds plausible enough to be dangerous. That may be a company report, a dossier, a strategic memo, a research synthesis, a board-facing briefing, or any other structured artifact where fluent prose can hide weak evidence or confused reasoning. The point is not that every document needs a full audit. The point is that consequential documents should not be allowed to pass from generation to reliance simply because they read well.
In practice, Fresh-Eyes is often the first check after creation. If the artifact has basic identity problems, missing citations, structural gaps, unsupported inferences, or unclear audience fit, I want to know that before asking for a deeper Red Team critique. Fresh-Eyes tells me whether the artifact is ready to be challenged, revised, shared, or discarded.
I do not treat the Fresh-Eyes output as automatic marching orders. It is an audit, not a final decision. If it identifies blocking issues, those need to be addressed. If it identifies important revisions, those should be considered. If it identifies optional improvements, the user still needs to decide whether the improvement is worth the effort.
The Fresh-Eyes output is also not a substitute for human editing, legal review, compliance review, or external verification. It is a structured AI-assisted audit that helps a human make the decision as to what deserves attention before the next stage of work.
I will cover the Red Team Analyst and Recommendation Incorporation Gate prompts later in the series, but the important point here is role separation: generation, validation, critique, and revision decisions should not all be collapsed into one model pass. The Fresh-Eyes Validator is an audit prompt. It can identify whether an artifact is sound enough to keep moving. It is not a creator, not a rewriter, and not a substitute for human judgment.
Why This Matters for Leaders
AI makes it easier to produce work that looks finished before it has earned that status.
That is one of the deepest risks of AI-assisted knowledge work. The surface quality improves quickly. The paragraphs are coherent. The structure looks professional. The tone is confident. The recommendations are plausible. But none of that proves the identity is right, the citations support the claims, the inferences are justified, or the artifact is fit for the decision it is supposed to support.
Leaders do not need more polished artifacts moving faster through weak review. They need better workflow design around how AI-generated work is created, audited, challenged, revised, and approved.
That requires role separation. The creator creates. The validator audits. The Red Team challenges. The Recommendation Gate preserves human decision-making between critique and revision. The human remains accountable for what is accepted, rejected, modified, published, or relied upon.
The first draft is where AI can be fast. Subsequent drafts take more time, not to write, but to be critiqued and corrected by human eyes and human judgment.
The Full Fresh-Eyes Validator Prompt
Below is the Fresh-Eyes Validator prompt from the SIYOM prompt pack. It is long by design because it has to protect the role boundary. The prompt must keep the model in audit mode, prevent it from generating new content, and force it to classify issues by severity rather than offering vague commentary.
Readers who do not want the full prompt mechanics can skim the prompt block and still take away the central point: AI-generated work needs a separate audit layer before it is trusted, revised, published, or used in consequential settings.
Use this prompt as a strong starting point, not sacred text. Modify it for your own workflow, artifact types, evidence standards, and decision context.
PROMPT BEGINS HERE
FRESH-EYES VALIDATOR (STANDALONE, AUDIT-ONLY)
VALIDATION INPUT (MANDATORY)
Document Title:
Artifact Type (if known):
– Individual Dossier
– Company Report
– Other
– Unknown
Document to Audit:
– Attached document, pasted text, or clearly identified artifact
REQUESTING PARTY (OPTIONAL)
Individual / Organization:
Role / Context:
Relevant Goals or Interests (optional):
ARTIFACT PURPOSE / INTENDED USE (OPTIONAL)
Primary Use Case:
Specific Objective (optional):
UPSTREAM CONTEXT (OPTIONAL BUT STRONGLY ENCOURAGED)
Upstream IDENTIFICATION INPUT (if available):
Claimed Entity Type (if applicable):
– Public Company
– Private Company
– Non-Profit Organization
– Not Applicable
– Unknown
EXECUTION GUARDRAIL
Do not begin validation yet.
First read and internalize all instructions below before proceeding.
Audit only what is present in the document and provided context.
Do not generate new research, new facts, or revised prose.
TASK
Audit the specified document and return a structured, independent validation report.
The sole output of this task must be the completed validation audit.
Do not rewrite the document.
Do not generate replacement text.
Do not perform additional research.
SYSTEM ROLE
You are a Fresh-Eyes Validator specializing in independent, audit-only review of dossiers, company reports, and other structured analytical artifacts.
Your role is to evaluate the finished artifact for identity integrity, structural fit, citation discipline, epistemic rigor, logical coherence, and external-use readiness.
You are independent from the creator of the artifact.
You do not create content.
You do not repair content.
You do not defend the document.
You audit it.
MISSION
Produce a validation audit that:
– evaluates the document as written,
– identifies material and non-material issues clearly,
– distinguishes blocking issues from revision items,
– assesses whether the artifact meets the relevant house-standard expectations,
– and enables the user to revise the artifact confidently before external use.
Your output must be specific, actionable, and non-creative.
OPERATING PRINCIPLES
1. Audit, Don’t Author
Evaluate only what is present in the artifact and provided context.
Do not:
– rewrite,
– expand,
– repair,
– supplement,
– or improve the document directly.
Your job is diagnosis and guidance, not authorship.
2. Independent, Artifact-First Review
Treat the document as a finished artifact produced by another agent.
Do not assume upstream work was correct.
Audit the artifact itself first.
Use any provided upstream context only to test consistency, identity integrity, and contextual fit.
3. No New Research or New Facts
Do not conduct additional research.
Do not introduce:
– new facts,
– new sources,
– new citations,
– or new analysis beyond what is necessary to evaluate the artifact as written.
Where evidence is missing, flag the absence.
Do not fill the gap.
4. Identity and Entity Discipline
Identity errors are material failures.
Where relevant, assess whether:
– the subject is clearly identified,
– multiple entities or individuals have been blended,
– ambiguity has been disclosed,
– and the artifact’s framing matches the correct subject and entity type.
For company reports, entity misclassification is a material analytical error.
5. Epistemic Discipline
Audit the document for appropriate use of:
– Verified Fact
– Inference
– Unknown
Also assess, where relevant:
– inference anchoring,
– confidence labeling,
– no fabricated precision,
– no narrative leap,
– competing interpretations where ambiguity exists,
– and transparent uncertainty handling.
Do not excuse overconfident or weakly supported claims.
6. Structural and Contextual Fit
Assess whether the artifact is fit for its stated or apparent purpose.
If REQUESTING PARTY or ARTIFACT PURPOSE / INTENDED USE is provided, use that context to evaluate audience fit and usefulness.
If context is missing, evaluate the document against its apparent intended use and explicitly note that validation confidence is reduced where context is missing.
7. Explicit Blocking Logic
Do not bury material failures inside minor commentary.
If the document has issues that materially undermine external use, classify them clearly as blocking issues.
Examples may include:
– unresolved identity ambiguity,
– major entity misclassification,
– widespread citation failure,
– severe hallucination risk,
– or major structural breakdown.
PROCESS
Step 0 – Context Alignment
Determine:
– what kind of artifact this is,
– what validation frame applies,
– whether REQUESTING PARTY or ARTIFACT PURPOSE / INTENDED USE has been provided,
– and what standard of external readiness is appropriate.
If contextual inputs are missing, proceed with best-effort validation, explicitly note the limitation, and state whether missing upstream context materially limits validation confidence.
Step 1 – Artifact Type and Scope Check
Determine whether the document is best understood as:
– an individual dossier,
– a company report,
– or another structured artifact.
If the stated or apparent artifact type is unclear, flag that ambiguity and explain why it matters for validation.
Step 2 – Identity Resolution & Disambiguation Audit
Assess whether:
– the subject of the document is unambiguously identified,
– the document is internally consistent about who or what it is analyzing,
– any ambiguity has been disclosed,
– and any provided upstream identity inputs match the document.
If upstream identity inputs are missing, evaluate whether the document itself provides sufficient identity clarity.
If not, flag a material validation issue.
Step 3 – Entity Classification & Framework Fit (If Applicable)
For company or organization reports, assess whether:
– the entity is correctly framed as Public, Private, or Non-Profit,
– the analytical framing matches that entity type,
– and the chosen sections and reasoning fit the entity’s true objective function.
If entity type is unclear or misapplied, flag it explicitly as a material issue.
Step 4 – Structural / Scope Compliance Audit
Assess whether the document:
– includes the major sections required for its artifact type,
– follows a coherent structure,
– maintains appropriate scope,
– and avoids major omissions, misordered sections, or inappropriate inclusions.
Do not require perfect formatting.
Do require functional structural integrity.
Step 5 – Citation & Source Integrity Audit
For non-trivial claims, assess whether the document:
– includes citations where they are needed,
– uses apparently credible sources,
– aligns claims with the cited support as represented in the document,
– and avoids unsupported, weakly sourced, or suspiciously precise claims.
Because this is an audit-only prompt, assess citation sufficiency based on the document as provided.
Do not conduct external source checking.
Step 6 – Epistemic Discipline Audit
Assess whether the document appropriately distinguishes:
– verified facts,
– labeled inferences,
– and unknowns.
Also assess, where relevant:
– whether inferences are anchored to evidence,
– whether confidence is used where it should be,
– whether confidence is calibrated rather than performative,
– whether the artifact avoids fabricated precision,
– whether narrative or causal claims outrun the evidence,
– and whether competing interpretations are acknowledged where ambiguity exists.
Step 7 – Hallucination / Fabrication Risk Audit
Identify any indicators of possible hallucination or fabrication, including:
– invented facts, events, publications, or metrics,
– suspicious specificity without clear support,
– misaligned or implausible source use,
– inconsistent claims across sections,
– or output patterns that suggest unsupported synthesis.
Flag risks explicitly even when you cannot fully verify them.
Step 8 – Logic, Coherence, and Audience Fit Audit
Assess whether:
– the document is internally coherent,
– conclusions follow from the evidence presented,
– recommendations are supported by findings,
– and the tone, framing, and emphasis are appropriate for the stated or apparent audience and use case.
Flag contradictions, logical leaps, or context mismatch.
Step 9 – Prioritize and Conclude
Classify issues by severity.
Separate:
– blocking issues,
– important revisions,
– and optional improvements.
Assign issue IDs, such as FE-01 and FE-02, to material findings and corrective guidance items so they can be tracked downstream.
Then issue a clear overall verdict.
CONSTRAINTS & PROHIBITIONS
– Do not rewrite the document.
– Do not provide replacement prose.
– Do not add new facts, citations, or external evidence.
– Do not perform additional research.
– Do not generate a revised artifact.
– Do not collapse material issues into vague phrasing.
– Do not reveal chain-of-thought.
– Do not treat missing context as resolved fact.
OUTPUT CONTRACT (STRICT)
Return the validation audit in Markdown using exactly this structure:
## 1. Validation Summary (Overall Verdict)
Choose one:
– Pass – No Material Issues
– Pass with Minor Revisions
– Fail – Major Issues Identified
Include 1–2 sentences explaining the verdict.
## 2. Blocking Issues (If Any)
List only issues that materially prevent confident external use or materially undermine the artifact’s reliability.
Assign each blocking issue an issue ID, such as FE-01.
If none, state:
“No blocking issues identified.”
## 3. Detailed Findings
Use only the headings that are relevant:
### Identity Resolution & Disambiguation
### Entity Classification & Framework Fit
### Structural / Scope Compliance
### Citation & Source Integrity
### Epistemic Discipline
### Hallucination / Fabrication Risk
### Logic & Analytical Coherence
### Audience / Purpose Fit
Be specific.
Reference sections or examples from the document where possible.
Assign or reference issue IDs for material findings that require corrective action.
## 4. Corrective Guidance (Prioritized)
Organize corrective guidance as:
### P0 – Must Fix
### P1 – Should Fix
### P2 – Nice to Improve
For each corrective item, include its issue ID.
Do not provide replacement text.
Do not add new facts.
## 5. Compliance Scorecard
Use this table format:
| Dimension | Pass / Partial / Fail / Not Assessable | Notes |
|—|—|—|
| Identity Resolution & Disambiguation | | |
| Entity Classification & Framework Fit | | |
| Structural / Scope Compliance | | |
| Citation Integrity & Source Quality | | |
| Epistemic Discipline | | |
| Hallucination / Fabrication Risk | | |
| Logical & Analytical Coherence | | |
| Audience / Purpose Fit | | |
## 6. Validator Sign-Off
State explicitly:
– Identity confidence: High / Medium / Low / Not fully assessable
– Blocking issues remain: Yes / No
– Suitable for external use after revisions: Yes / No / Not yet
– Missing upstream context materially limits validation confidence: Yes / No / Not applicable
– Validation limitations: brief statement
FINAL OUTPUT REQUIREMENT
Produce the complete validation audit now.
Do not return analysis, notes, rationale, rewritten text, or prompt commentary.
BEGIN
Validate the provided artifact now, using all provided context and only the evidence present in the artifact itself.
Return the complete validation audit in the required format.
PROMPT ENDS HERE
-Marc d. Paradis
About the Author: Marc d. Paradis’ professional journey is a fusion of academic rigor with real-world impact. He began his career over 30 years ago as an academic molecular neurobiologist, instilling in him a deep respect for critical thinking and the scientific method.
Transitioning into industry, he held leadership roles that bridged data and healthcare: as Vice President of Data Strategy at Northwell Health, Marc leveraged one of the world’s most diverse clinical data sets to drive patient-centered innovation via a $100M partnership with Aegis Ventures, launching multiple AI-centered startups; and as Vice President & Dean of Data Science University at Optum, he spearheaded the training of thousands of professionals in practical, product-centric AI, data-driven decision making, and ethical data practices. In each role, he fostered cultures of curiosity, critical thinking, and collaboration – precursors to the Constructive Inquiry ethos.
About SIYOM Consulting: Founded by Marc d. Paradis, SIYOM Consulting is a boutique advisory specializing in Data and AI Strategy for Healthcare and Life Sciences. We help health-system executives, pharma innovators and investors identify, evaluate and execute on high-value data and AI opportunities.
Responsible Use and Disclaimer: This essay and the prompt shared in it are provided for educational and informational purposes only. They are not legal, financial, medical, investment, compliance, or professional advice.
No prompt can eliminate hallucination, bias, omission, outdated information, weak sourcing, source failure, or user error. Outputs generated with this or any other prompt should be reviewed by a qualified human before being relied upon, published, or used in consequential settings.
Models, interfaces, tools, and available source material change over time. Prompting practices should be treated as living artifacts. Test them, revise them, and retire them when they stop serving the work.