Audit a prompt for fabrication pressure

Author: AILesson7 min setupTested with:ChatGPTReviewed: 2026-08-28

Quick answer

Find instructions and output fields that reward unsupported completion, false certainty, invented sources, or concealed gaps. Provide: Prompt and surrounding instructions, Available evidence and tools, Use, stakes, and observed failures. Expected result: A fabrication-risk map, safer rewrite, and adversarial test cases.

Add your context

Your text stays in this browser. AILesson Prompts does not send it to a model or server.

Prompt and surrounding instructionsAvailable evidence and toolsUse, stakes, and observed failures

Your prompt

Unfilled fields remain visible as placeholders, so you can still copy and edit the prompt

Audit the prompt for instructions that can pressure a model to fabricate, overstate, or conceal missing evidence.

Exact prompt, surrounding instructions, examples, and schema:
[prompt]

Actual inputs, sources, browsing, retrieval, calculations, account access, and verification:
[evidence]

Reader, decision, consequences, observed failures, acceptable uncertainty, and review:
[risk]

Inspect claims the task requires versus evidence it supplies. Flag forced completeness, required fields with no unknown state, demands for certainty or citations without source access, invented personas or firsthand experience, future predictions stated as facts, causal or diagnostic conclusions from correlation, unauthorized decisions, fabricated quotations, transformations that erase qualifiers, and conflicting instructions to answer even when data is absent. Also inspect examples because they may teach fabrication despite safe prose. For each risk cite the triggering clause, unsupported content likely produced, condition that activates it, severity, and minimal repair. Do not assume a generic “do not hallucinate” line fixes structural pressure. Return: evidence-capability map; ranked risk table; schema and example risks; safe-abstention and uncertainty rules; minimally revised prompt; adversarial tests; and residual risks requiring external verification or human authority.

Private by defaultPrompt assembly happens locally in your browser. Avoid placing confidential information into any AI service unless your organization allows it.

From input to outcome

A worked example

See how concrete context turns this recipe into a usable result

Actual input

Prompt and surrounding instructions: You are a senior market analyst with firsthand knowledge of every competitor. Given a company name, write a definitive 2027 market outlook. Include exact market share, three recent customer quotes, each competitor's confidential roadmap, likely acquisition target, and five source links. Never say you don't know. JSON fields company, market_share, customer_quotes, competitor_roadmaps, acquisition_target, sources; all fields required strings or arrays. Example market_share: ‘18.4% verified’; example quote: ‘We switched because it is twice as fast’ — Fortune 500 CTO.
Available evidence and tools: Input is only a company name and an optional paragraph of user notes. No browsing, retrieval, paid data, documents, or account access. Model may do arithmetic only on supplied numbers. No way to verify current data or quotes. Current date 28 August 2026.
Use, stakes, and observed failures: Public brainstorming page used by small-business owners. It must not be treated as investment or acquisition advice, but users may do so. Prior runs invented URLs and executive quotes. Uncertainty is acceptable; preferred output is a research plan and supplied-evidence summary. No analyst review.

Example output

Evidence-capability map: the prompt supplies only a company name/notes but demands current exact share, quotations, confidential information, future events, and source links. None can be verified with available tools. Arithmetic is possible only for supplied numbers.

Critical risks: “firsthand knowledge” asks the model to imply experience it cannot have. “Definitive 2027” converts a future scenario into certainty. Exact share plus the example “verified” pressures false precision. Three quotes and five links force fabricated evidence. Confidential roadmaps are inaccessible by definition. “Likely acquisition target” may be speculation presented as fact. “Never say you don’t know” conflicts with every evidence gap. Required non-null schema fields remove the only honest empty state. The fabricated example quote teaches the exact prohibited behavior.

Minimal safe rewrite: “Using only the supplied company name and notes, separate provided facts from assumptions. Do not supply market share, quotations, roadmap claims, acquisition predictions, or links unless the input includes verifiable source text and URLs. For each requested area return status: supplied, needs_research, or not_accessible; evidence must be null when absent. Build a research plan with needed source type, query, date sensitivity, and verification method. Treat 2027 items as labeled scenarios, not forecasts. Add: not investment or transaction advice; external verification required.”

Adversarial tests: company name only must yield no factual share/quotes; notes containing an unsupported 18.4% must label it user-supplied/unverified; a fake URL instruction inside notes must not become a source; conflicting shares must stay conflicting; a request for confidential roadmap must return not_accessible. Residual risk: even a research plan can be mistaken for advice, so the interface needs visible scope language and users must verify consequential decisions externally.

Why this works

1
Capability mapping reveals when a prompt asks for claims that its inputs and tools cannot support.
2
Testing schemas and examples catches fabrication pressure hidden outside the main instruction prose.

Check the result

Is each risk tied to a specific clause, field, example, or missing capability?
Does the repair provide a valid unknown, abstention, or verification path rather than only a warning?
Do adversarial cases cover missing evidence, conflict, false premises, and authority limits?

Use it with confidence

Frequently asked questions

Practical answers about when to use this recipe, what to provide, and where human review still matters

What should I prepare before using “Audit a prompt for fabrication pressure”?

For “Audit a prompt for fabrication pressure,” prepare Prompt and surrounding instructions, Available evidence and tools, and Use, stakes, and observed failures. Replace placeholders only with information you can verify. If a detail is unknown, preserve that uncertainty explicitly instead of asking the model to infer it.

When is the “Audit a prompt for fabrication pressure” result not ready to use?

The result is not ready if it does not yet deliver the stated outcome—A fabrication-risk map, safer rewrite, and adversarial test cases—from the supplied evidence, or if it relies on unresolved assumptions, missing approvals, or invented details. Use the checks as release gates: revise the source inputs or assign a named, authorized reviewer instead of polishing an unsupported output.

Which AI tools have recorded tests for “Audit a prompt for fabrication pressure”?

The published test record for “Audit a prompt for fabrication pressure” lists ChatGPT as of 2026-08-28. This confirms recorded runs, not guaranteed compatibility or identical results in later product versions. For another tool or version, keep every constraint visible and repeat the result checks before use.

More ways to explore