25 ChatGPT-5.5 Prompts for Source-Controlled Deep Research: Plan Review, Domain Filters, Live Steering, Citation Checks, and Decision Artifacts

How to Use ChatGPT-5.5 as the Control Layer for Deep Research

These prompts are designed for teams that need research outputs they can defend: market scans with explicit source boundaries, vendor comparisons with evidence grades, policy briefs with unresolved questions, executive memos with citation checks, and implementation plans that separate findings from recommendations. The practical role of ChatGPT-5.5 in this workflow is to structure the brief, test the research plan, express source controls, draft steering messages, apply verification rubrics, and shape decision artifacts. It should not be described as the universal underlying model that powers every Deep Research task across every ChatGPT account, workspace, or surface.

OpenAI documents Deep Research as a workflow that starts with a defined outcome and report structure, uses context and sources supplied or authorized for the task, may ask clarifying questions, can be monitored and steered while it runs, and returns a structured report with citations or source links. That makes prompt quality operationally important: the first prompt should not merely ask for “research”; it should define the decision, audience, timeframe, source perimeter, evidence hierarchy, artifact format, and verification checkpoints before the research run becomes expensive, broad, or difficult to audit.

The model naming boundary also matters. The official GPT-5.5 model documentation is the relevant catalog reference for GPT-5.5 usage, and no official gpt-5.5-mini identifier is documented in the supplied source set. Do not build internal prompt libraries, routing tables, cost calculators, or governance policies around an undocumented mini identifier. If your application or workspace exposes different model options, treat those as environment-specific availability details rather than facts that can be generalized to every reader or deployment.

The article covers advanced prompt engineering techniques for ChatGPT, Claude, and Codex in 2026, including chain-of-thought, tree-of-thought, and agent prompting. The complete Advanced Prompt Engineering for ChatGPT, Claude, and Codex in 2026 article provides the destination-specific detail for this section’s ChatGPT-5.5 Prompt Engineering decision because it is the strongest broad prompt-engineering companion for a ChatGPT-5.5 prompt collection, giving readers technique-level context for constructing more reliable prompts.

The prompts in this article use ChatGPT-5.5 as a disciplined reasoning, planning, and drafting partner around Deep Research. In practice, that means you can ask it to convert a vague executive request into a research contract, challenge an overbroad plan, propose allowed and excluded domains, draft a mid-run steering message, build a claim-to-source matrix, identify unsupported claims, and produce a decision memo that tells human reviewers exactly what must be checked before circulation.

Operational warning: Citations improve traceability, but OpenAI’s Deep Research guidance says users should verify underlying sources, especially for decisions, external publishing, or high-stakes topics. A cited sentence is not automatically correct, current, complete, licensed for reuse, or appropriate for legal, medical, financial, security, or compliance conclusions.

The Source-Controlled Research Pattern

Source-controlled research is a prompting pattern that treats sources as part of the task specification rather than as an afterthought. Instead of asking for a general answer, you tell the model which sources are permitted, which sources are preferred, which sources are excluded, what date range matters, what counts as primary evidence, and what claims require human or expert review. This reduces ambiguity and makes the final artifact easier to inspect because the user can compare each major conclusion against the allowed evidence base.

OpenAI states that in Chat, users can review and modify a proposed Deep Research plan and restrict research to specified websites or prioritize selected sites while still allowing broader web search. In Work or Codex, users can add instructions to steer or revise scope. The prompts below turn those capabilities into repeatable language: plan-review prompts before launch, domain-filter prompts for scope control, live-steering prompts during research, and verification prompts after the report is returned.

Permission boundaries remain central. OpenAI’s Deep Research documentation says starting a Deep Research task does not grant access to additional files or apps; workspace, role, session-tool, provider, and connected-account permissions continue to apply. Deep Research in Chat can use the public web and uploaded files by default, and supported connected apps when available, but connected app access depends on what is configured and authorized. In Work and Codex, requested output to a connected app depends on supported capabilities and permissions.

That distinction prevents a common governance mistake: a prompt can request a spreadsheet, document, presentation, or other artifact, but it cannot guarantee that the tool exists, that the connected app supports the requested action, that the current account has permission, or that the generated file opens successfully. OpenAI’s guidance requires users to verify that output files and links exist and open as expected. The prompts therefore include artifact checks instead of assuming successful delivery.

The article is a practitioner playbook for building reliable, verifiable, and scalable deep research AI systems using Valyu MCP, real-world data, RAG, verification, governance, and observability. The complete The Valyu Deep Research Playbook — Connecting Codex to Real-World Data article provides the destination-specific detail for this section’s Deep Research Best Practices decision because its focus on verification, governance, and scalable deep research systems closely matches the current article’s emphasis on controlled research workflows and citation checks.

The Reusable Prompt Contract

Every prompt in this masterclass follows the same reusable contract so teams can store, review, and adapt the templates without losing the control logic. The contract is intentionally explicit because Deep Research tasks often fail when the user leaves the decision, audience, source rules, and evidence thresholds implicit. Treat the contract as the minimum viable specification before launching or steering a research run.

Contract field	What to specify	Why it matters
Decision or task outcome	The decision, recommendation, artifact, or next action the research must support.	Prevents the report from becoming a general literature review with no operational conclusion.
Audience	The reader’s role, technical depth, risk tolerance, and expected use of the output.	Changes the level of detail, terminology, evidence threshold, and artifact format.
Scope and timeframe	Geography, market segment, product area, regulation, competitor set, or date range.	Constrains search and helps reviewers detect stale or out-of-scope evidence.
Permitted and preferred sources	Official sites, primary documents, uploaded files, named connected sources, or approved domains.	Turns source selection into an auditable instruction rather than a hidden model choice.
Excluded sources	Domains, content types, regions, unauthoritative summaries, or materials without reuse rights.	Reduces contamination from irrelevant, promotional, outdated, or legally risky material.
Evidence hierarchy	How to rank primary sources, official documentation, peer-reviewed work, expert commentary, news, and secondary summaries.	Helps resolve contradictions and makes weak support visible instead of blending it into confident prose.
Output format	Memo, table, spreadsheet-ready matrix, slide brief, checklist, risk register, or implementation plan.	Aligns the research run with a usable artifact rather than requiring a second reconstruction step.
Verification checkpoint	The human review action required before relying on or sharing the output.	Preserves accountability where citations, permissions, sensitivity, and expert judgment matter.

Each copy-paste prompt also requires the model to distinguish evidence, inference, uncertainty, and unsupported claims. This is not cosmetic formatting. Evidence is what a source directly supports; inference is the model’s or analyst’s reasoned interpretation; uncertainty identifies missing or conflicting information; unsupported claims are statements that should be removed, reframed, or queued for additional research. This separation gives reviewers a faster way to find the parts of a report that should not be forwarded without inspection.

For external sharing, OpenAI Academy advises users to recheck citations, remove sensitive information, confirm rights and permissions for included material, avoid long copied passages, and retain citations. The prompt contract therefore treats publication as a separate review state, not as the automatic end of a Deep Research task. A polished document can still contain sensitive internal facts, overlong quotations, unsupported synthesis, or material that is not cleared for customer, investor, regulator, or public distribution.

How to Adapt the 25 Templates Without Weakening Controls

When adapting a template, change the business nouns first and the control clauses last. For example, replace “vendor scorecard for customer-support automation” with “vendor scorecard for identity-governance tooling,” but keep the instructions that require source classification, citation preservation, uncertainty labeling, and human review. Removing those clauses often saves a few lines while increasing the risk that a confident artifact hides weak evidence.

Use narrower source filters when the decision depends on official policy, product documentation, legal obligations, security posture, or regulated claims. Use broader discovery only when the goal is landscape mapping, hypothesis generation, or identifying candidate sources for later verification. A market-scan prompt can tolerate more exploratory evidence if it labels uncertainty clearly; a compliance or procurement prompt should privilege official documentation, contracts, audited materials, and counsel-reviewed sources where applicable.

Live steering should be concise and corrective. If the research plan drifts, interrupt with a short message that states what to stop doing, what to prioritize, which sources to add or exclude, and how the final artifact should change. The best steering prompts do not restart the entire task; they modify the active plan, preserve useful work already completed, and require the final report to disclose any scope changes that affected the evidence base.

After the report is generated, run a citation audit before turning it into a decision artifact. Ask for a claim-to-source matrix, then inspect the highest-impact claims manually against the cited sources. Pay special attention to comparative statements, numerical claims, vendor claims, current availability, legal interpretations, financial implications, medical or safety advice, and any recommendation that would trigger spending, customer communication, employee action, or operational change.

The 25 prompts that follow are sequential, but they are not a mandatory linear workflow. Use the framing and source-control prompts before launching research, the plan-review and live-steering prompts while the task runs, the citation and contradiction prompts after the report returns, and the decision-artifact prompts when you need a memo, scorecard, slide brief, spreadsheet, or implementation plan that a human owner can approve.

Prompts 1–9: Lock the Research Question, Sources, Permissions, Plan, and Assumptions Before Deep Research Runs

Use these first nine templates before or at the start of a Deep Research task. They are designed to make ChatGPT-5.5 useful as a planning, scoping, and review layer without assuming that ChatGPT-5.5 is the underlying Deep Research model on every account, workspace, or surface. OpenAI’s Deep Research guidance emphasizes defining the outcome, report structure, context, sources, timeframe, and evidence boundaries, then reviewing the plan, steering the work while it runs, and validating citations before reuse.

Prompt 1: Decision-Question Framing

Purpose: Convert a broad topic into a decision-grade research question. This prevents Deep Research from producing a general explainer when the user actually needs a recommendation, go/no-go assessment, vendor shortlist, risk register, or board-ready decision memo.

Copy-paste prompt

You are helping me prepare a source-controlled Deep Research task. Convert the topic below into one primary decision question and up to five sub-questions.

Topic:
{{topic}}

Decision context:
{{decision_context}}

Constraints:
{{business_constraints}}

For each proposed question, state what decision it supports, what evidence would answer it, and what evidence would be insufficient. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. Do not begin broad research yet. First produce a question frame I can approve or revise.

Required inputs:

• The topic or problem statement.
• The decision that must be made after research.
• Known constraints such as budget, geography, regulation, customer segment, or deadline.

Expected output: A primary decision question, supporting sub-questions, evidence requirements, and a warning list showing what kinds of sources or claims would not be enough to support a decision.

Verification checkpoint: Confirm that the question asks for a decision, not merely a summary. If the answer could be “interesting background,” rewrite it until it can drive an explicit action.

Prompt 2: Audience and Artifact Contract

Purpose: Define who will read the final output and what artifact they need. OpenAI documents that Deep Research can produce structured reports with citations and, when the needed tools are available, may support editable documents, presentations, spreadsheets, or other artifacts; users still must verify that generated files or links exist and open successfully.

Copy-paste prompt

Before research begins, create an audience and artifact contract for this Deep Research task.

Audience:
{{audience}}

Intended use:
{{intended_use}}

Preferred artifact:
{{report_document_presentation_spreadsheet_or_other}}

Required sections:
{{required_sections}}

Tone and depth:
{{tone_depth}}

Define the artifact contract as: audience, decision supported, required format, required sections, citation expectations, evidence standard, items excluded from scope, and final review owner. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. Do not assume that every output format or connected-app action is available; mark tool-dependent items separately.

Required inputs:

• Reader group, such as executives, counsel, engineering leaders, procurement, or research teams.
• Final artifact type and required sections.
• Decision or workflow the artifact will support.

Expected output: A concise artifact specification that can be pasted into a Deep Research request, including format, section order, citation rules, and review responsibilities.

Verification checkpoint: Check whether the artifact contract names an accountable reviewer. If the output may be externally shared, add review for sensitive information, rights, permissions, and citation support.

The article provides advanced reusable prompt patterns for research workflows on GPT-5 Pro and GPT-5.4, including literature triage, study extraction, argument mapping, and synthesis techniques. The complete Advanced Prompt Patterns for research: Working Examples for GPT-5 Pro and GPT-5.4 article provides the destination-specific detail for this section’s Research Prompt Templates decision because it directly supports the marker by offering research-specific prompt patterns rather than generic prompt templates.

Prompt 3: Timeframe and Recency Boundary

Purpose: Set the historical and recency window. A market scan may require the last 12 months; a regulatory analysis may require current obligations plus pending changes; a technical review may need both foundational papers and current implementation evidence.

Copy-paste prompt

Define the timeframe for a Deep Research task and explain how recency should affect evidence weight.

Research question:
{{research_question}}

Decision deadline:
{{decision_deadline}}

Required historical window:
{{historical_window}}

Recency needs:
{{recency_needs}}

Create a timeframe policy with: default date range, must-include historical sources, must-prioritize recent sources, stale-source warning rules, and how to handle sources with no publication date. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. If currentness is essential, state that citations must be checked against the underlying source before any decision or external sharing.

Required inputs:

• Decision deadline or publication date.
• Research domain, because acceptable freshness differs by field.
• Known events, product releases, policy changes, or market dates that anchor the review.

Expected output: A timeframe policy that tells Deep Research which sources are timely, which are useful only as background, and which should be flagged as potentially stale.

Verification checkpoint: Inspect whether the prompt defines both “include” and “downgrade” rules. If no source date is available, the task should treat the claim as lower-confidence unless corroborated elsewhere.

Prompt 4: Evidence Hierarchy

Purpose: Rank source types before the model starts collecting evidence. This reduces the risk that a blog post, summary, or vendor claim receives the same weight as primary documentation, audited filings, peer-reviewed research, official policy, or direct user-provided materials.

Copy-paste prompt

Create an evidence hierarchy for this Deep Research task.

Research question:
{{research_question}}

Domain:
{{domain}}

Preferred evidence types:
{{preferred_evidence_types}}

Known weak evidence types:
{{weak_evidence_types}}

Produce a ranked evidence hierarchy from strongest to weakest. Include rules for resolving conflicts, handling vendor-authored claims, using news or commentary, and treating uncited claims. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. When evidence conflicts, do not average the claims; explain which source should be weighted more and why.

Required inputs:

• The decision question and domain.
• Primary-source categories that should be preferred.
• Source categories that may be biased, outdated, incomplete, or secondary.

Expected output: A weighted evidence ladder, conflict-resolution rules, and a citation-handling policy for weak or unsupported claims.

Verification checkpoint: Confirm that the hierarchy fits the domain. For legal, medical, financial, security, or regulated topics, require expert review and do not allow the final artifact to present conclusions as professional advice.

Prompt 5: Allowed Domains and Source Whitelist

Purpose: Define allowed web domains or source collections. OpenAI states that Deep Research in Chat can let users restrict research to specified websites or prioritize selected sites while allowing broader web search, and that source access depends on the current task and available permissions.

Copy-paste prompt

Prepare an allowed-source policy for Deep Research.

Research question:
{{research_question}}

Allowed domains or source collections:
{{allowed_domains_or_sources}}

Should full-web search be allowed after checking these sources?
{{yes_no_and_conditions}}

Priority order:
{{priority_order}}

Create a source policy that says: search these sources first, use full-web search only under these conditions, do not treat allowed sources as automatically correct, and preserve citations for every material claim. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. If an allowed source cannot answer a sub-question, mark the gap rather than filling it from an unapproved source.

Required inputs:

• Allowed domains, internal files, or approved source collections.
• Whether broader web search is permitted and under what conditions.
• Priority order for official, internal, academic, industry, or vendor sources.

Expected output: A source whitelist, search order, fallback rule, and gap-reporting instruction that can be reused in the main Deep Research prompt.

Verification checkpoint: Verify that allowed does not mean authoritative. The final report still needs citation checks, contradiction review, and human validation for high-stakes decisions.

The article presents research-grade ChatGPT prompts designed to minimize hallucinated citations and force calibrated uncertainty across GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro. The complete Best ChatGPT Prompts for research article provides the destination-specific detail for this section’s Source Constrained Research decision because it is the best match for source-constrained research because it explicitly addresses citation reliability, hallucination reduction, and uncertainty handling in research prompts.

Prompt 6: Excluded Sources and Disallowed Evidence

Purpose: Block sources that should not influence the result. Exclusions may include competitor marketing pages, unsourced AI-generated summaries, stale commentary, forums, non-public materials without permission, or documents outside the approved confidentiality boundary.

Copy-paste prompt

Define excluded sources and disallowed evidence for this Deep Research task.

Research question:
{{research_question}}

Sources to exclude:
{{excluded_sources}}

Evidence types to avoid:
{{disallowed_evidence_types}}

Reason for exclusions:
{{reasons}}

Create an exclusion policy that names sources, categories, and evidence patterns that must not be used. Explain whether they may be mentioned as context but not relied upon. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. If excluded material appears in search results, ignore it for conclusions and list it only in an excluded-source note.

Required inputs:

• Specific domains, publications, files, or source categories to exclude.
• Reasons for exclusion, such as bias, rights concerns, confidentiality, staleness, or lack of citations.
• Whether excluded sources may be noted as background but not used as evidence.

Expected output: A disallowed-source policy with handling rules for accidental discovery, ambiguous sources, and uncited claims.

Verification checkpoint: Confirm that the exclusions do not remove necessary primary evidence. If they do, revise the decision question or add an approved substitute source.

Prompt 7: Connected-App Permission Inventory

Purpose: Inventory which files, apps, and connected sources are actually authorized. OpenAI states that starting Deep Research does not grant additional file or app access; workspace, role, session-tool, provider, and connected-account permissions continue to apply. In Chat, connected apps are used for supported read actions in research rather than arbitrary write actions.

Copy-paste prompt

Create a connected-source permission inventory before Deep Research starts.

Research question:
{{research_question}}

Uploaded files:
{{uploaded_files}}

Connected apps or internal sources expected:
{{connected_apps_or_sources}}

Known permission boundaries:
{{permission_boundaries}}

Produce an inventory with columns for source name, owner, access status, permitted use, read/write expectation, sensitivity level, and unresolved permission questions. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. Do not assume access to any file, folder, app, repository, workspace, or provider account that is not explicitly authorized in this task.

Required inputs:

• Files uploaded for the task.
• Connected apps or internal systems expected to be used.
• Known workspace, role, account, or provider permission constraints.

Expected output: A permission table identifying authorized sources, missing access, sensitivity concerns, and questions to resolve before the research run.

Verification checkpoint: Check whether any planned source requires access the user does not have. If so, either obtain proper access outside the prompt or remove that source from scope.

Prompt 8: Plan-Review Red Team

Purpose: Stress-test the proposed research plan before approving it. OpenAI’s Deep Research flow allows users to review and modify a proposed plan in Chat and to steer or revise scope in Work or Codex, so the plan review is the best time to catch source gaps, ambiguous scope, and unsupported decision criteria.

Copy-paste prompt

Act as a red-team reviewer for this proposed Deep Research plan.

Proposed plan:
{{proposed_plan}}

Decision question:
{{decision_question}}

Source policy:
{{source_policy}}

Artifact contract:
{{artifact_contract}}

Review the plan for missing sources, overbroad scope, biased evidence, unclear decision criteria, unsupported assumptions, permission problems, stale timeframe, and citation risks. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. Return: approve, approve with changes, or reject and rewrite. Do not proceed to research until the plan is corrected.

Required inputs:

• The proposed Deep Research plan.
• The decision question, source policy, and artifact contract.
• Any known constraints, such as deadline, geography, audience, or excluded sources.

Expected output: A red-team review with approval status, required edits, unresolved risks, and a rewritten plan if needed.

Verification checkpoint: Do not approve a plan that lacks source boundaries, decision criteria, or citation expectations. A polished plan can still be unsafe if it omits permission checks or assumes facts not in evidence.

Prompt 9: Assumptions Register

Purpose: Capture assumptions before they become hidden conclusions. This is essential when research depends on market size estimates, regulatory interpretation, customer intent, roadmap timing, internal capacity, vendor claims, or incomplete source access.

Copy-paste prompt

Create an assumptions register for this Deep Research task.

Decision question:
{{decision_question}}

Known facts:
{{known_facts}}

Unknowns:
{{unknowns}}

Initial assumptions:
{{initial_assumptions}}

Build an assumptions register with columns for assumption, source or basis, confidence, impact if wrong, validation method, owner, and status. For every material claim, label it as Evidence, Inference, Uncertainty, or Unsupported; preserve citations/source links exactly where available; and identify items requiring human or expert review. Separate facts from assumptions and do not let assumptions appear as conclusions in the final artifact unless validated.

Required inputs:

• Known facts already accepted by the team.
• Unknowns that research should resolve or narrow.
• Initial assumptions that may influence the decision.

Expected output: A structured assumptions register that travels with the research plan and final artifact, including confidence levels and validation owners.

Verification checkpoint: Compare the final report against the register. Any assumption still unvalidated should remain labeled as an assumption, uncertainty, or decision risk rather than being promoted to a factual finding.

Prompts 10–18: Steer the Run, Stress-Test the Evidence, and Convert Research Into Decision Artifacts

The next nine templates are designed for the middle and late stages of source-controlled Deep Research: correcting scope while a task is running, forcing contradiction checks, finding missing evidence, mapping claims to citations, and turning findings into artifacts that a team can review. OpenAI documents that users can monitor progress, interrupt or steer Deep Research, and receive structured reports with citations or source links, but it also warns that citations improve traceability rather than guaranteeing correctness. Use ChatGPT-5.5 here as a prompt-design and review layer; do not assume it is the underlying Deep Research model for every account, workspace, or surface.

Prompt 10: Live Scope Correction

Purpose

Use this when the Deep Research task is drifting into irrelevant sources, over-broad comparisons, outdated material, or unsupported speculation. The goal is to interrupt the run with a precise correction while preserving useful work already completed.

Copy-paste prompt

You are helping me steer an active Deep Research task without restarting from scratch.

Current research objective:
[PASTE ORIGINAL OBJECTIVE]

Observed drift or problem:
[DESCRIBE WHAT IS GOING WRONG: wrong geography, wrong customer segment, outdated sources, weak citations, too much background, missing competitor class, etc.]

Corrected scope:
[STATE THE NEW BOUNDARY]

Instructions:
1. Preserve any already-found evidence that still fits the corrected scope.
2. Stop expanding into topics, geographies, source types, or time periods outside the corrected scope.
3. Mark each retained point as Evidence, Inference, Uncertainty, or Unsupported claim.
4. Preserve citations and source links for every evidence-backed point.
5. Flag any citation that no longer supports the corrected scope.
6. Identify items requiring human or expert review before they are used in a decision artifact.
7. Return a revised research plan with: keep, discard, investigate next, and unresolved questions.

Required inputs

• The original objective or plan.
• A short description of the drift you observed.
• The corrected audience, geography, timeframe, source boundary, or decision question.

Expected output

You should receive a revised plan that separates retained evidence from discarded material and shows what the research agent should investigate next. A strong answer will not merely apologize for drift; it will tell you exactly which sources, claims, and subtopics remain valid under the corrected boundary.

Verification checkpoint

Before accepting the redirected run, verify that the corrected scope is narrower than the original and that the retained citations still support the retained claims. If the task relies on uploaded files or connected apps, remember that Deep Research can only use sources available and authorized in the current session, workspace, role, and connected account.

Prompt 11: Contradiction Search

Purpose

Use this after an initial report looks too clean. The prompt asks Deep Research to actively search for conflicting evidence, dissenting interpretations, changed facts, and source incentives that could weaken a decision.

Copy-paste prompt

Run a contradiction search against the current research findings.

Decision or claim set to test:
[PASTE SUMMARY, CLAIMS, OR DRAFT RECOMMENDATION]

Scope limits:
[GEOGRAPHY, TIMEFRAME, INDUSTRY, CUSTOMER TYPE, SOURCE RULES]

Instructions:
1. Search for credible sources that contradict, qualify, narrow, or update the current findings.
2. Separate Evidence, Inference, Uncertainty, and Unsupported claims.
3. Preserve citations and source links for both supporting and contradicting evidence.
4. Identify whether each contradiction is factual, definitional, methodological, temporal, geographic, commercial, or opinion-based.
5. Do not treat a citation as decisive merely because it exists; explain what the source actually supports.
6. Identify items requiring human or expert review, including legal, financial, medical, security, compliance, or high-stakes operational questions.
7. Produce a contradiction table and a revised confidence rating for each major claim.

Required inputs

• The finding, memo, plan, or recommendation to test.
• Source boundaries and date limits.
• The kind of contradiction that matters most, such as pricing conflict, regulatory conflict, technical feasibility, or customer demand.

Expected output

The output should be a table of contested claims, contradicting evidence, supporting evidence, and a confidence adjustment. Useful contradiction searches often reveal that two sources are both accurate but refer to different years, markets, definitions, product editions, or regulatory contexts.

Verification checkpoint

Open the sources behind the most important contradictions and confirm that the quoted or summarized material appears in the cited source. Treat unresolved conflicts as decision risks, not as model failures to be smoothed over.

Prompt 12: Missing-Evidence Hunt

Purpose

Use this when the research report is persuasive but may be incomplete. The template forces a gap analysis so the final artifact does not hide missing market data, absent primary sources, unverified implementation assumptions, or unsupported competitive claims.

Copy-paste prompt

Audit the current research for missing evidence.

Current research summary:
[PASTE SUMMARY OR REPORT EXCERPT]

Decision this research is meant to support:
[STATE DECISION]

Evidence standard:
[DEFINE ACCEPTABLE SOURCE TYPES: official docs, filings, peer-reviewed papers, interviews, first-party product docs, analyst reports, customer tickets, etc.]

Instructions:
1. Identify the strongest claims that lack adequate evidence.
2. Classify each item as Evidence, Inference, Uncertainty, or Unsupported claim.
3. Preserve citations and source links for existing evidence.
4. For missing evidence, specify the exact source type or document needed.
5. Distinguish unavailable evidence from evidence that was not yet searched.
6. Identify items requiring human or expert review before use.
7. Return a missing-evidence backlog with priority, decision impact, likely source location, and recommended next search query or internal owner.

Required inputs

• The current report or draft findings.
• The decision that depends on the evidence.
• Your minimum evidence standard.

Expected output

Expect a backlog rather than a rewritten report. The best result ranks gaps by decision impact, such as “blocks launch recommendation,” “affects vendor shortlist,” or “needs counsel review before external publication.”

Verification checkpoint

Do not ask the model to fill gaps with assumptions. If a missing item cannot be found in authorized web, uploaded, or connected sources, record it as unavailable and assign a human owner.

This evidence-safe healthcare research tutorial shows how to separate public sources from connected organizational context, document permissions, and validate evidence boundaries, offering a concrete domain example for the general evidence-audit prompts in this masterclass. The complete How to Use Healthcare Public Data in ChatGPT and Codex: Complete Setup and Evidence-Safe Research Tutorial article provides the destination-specific detail for this section’s AI Evidence Audit decision because the target is an operational evidence-safe research tutorial and is a more exact audit companion than a broad healthcare data comparison.

Prompt 13: Claim-to-Source Matrix

Purpose

Use this to convert a narrative report into an auditable evidence table. The matrix is especially useful before sending a report to executives, counsel, security reviewers, procurement, or a customer-facing team.

Copy-paste prompt

Create a claim-to-source matrix from the current research.

Source material:
[PASTE REPORT, NOTES, OR RESEARCH OUTPUT]

Instructions:
1. Extract every decision-relevant claim.
2. For each claim, classify it as Evidence, Inference, Uncertainty, or Unsupported claim.
3. Preserve citations and source links exactly as provided.
4. Explain what each citation supports and what it does not support.
5. Identify claims that rely on multiple sources, stale sources, source interpretation, or no source.
6. Mark any claim requiring human or expert review before publication or decision use.
7. Return a table with columns: Claim, Classification, Citation or source link, Source type, What the source supports, Limits of support, Confidence, Review owner.

Required inputs

• A report, memo, slide draft, spreadsheet summary, or research notes.
• Any required source categories, such as official documentation, customer interviews, or regulator publications.
• The intended use of the artifact.

Expected output

The output should make weak support visible. A claim like “customers prefer vendor A” may become three rows: interview evidence from a small sample, inferred buying preference, and an unsupported market-wide generalization that should be removed or narrowed.

Verification checkpoint

Randomly sample high-impact rows and open their citations. If the matrix says a source supports a claim, confirm that the source contains the fact, not merely a related concept.

Prompt 14: Market Landscape

Purpose

Use this to build a source-constrained market map without overclaiming market size, maturity, or competitive position. It is suitable for founders, product leaders, investors, and strategy teams that need a structured view before deeper validation.

Copy-paste prompt

Conduct a source-controlled market landscape review.

Market or category:
[DEFINE MARKET]

Decision to support:
[ENTER DECISION: enter market, prioritize segment, build partnership list, prepare diligence, etc.]

Scope:
[GEOGRAPHY, CUSTOMER SEGMENT, TIMEFRAME, INCLUDED/EXCLUDED VENDOR TYPES]

Permitted or preferred sources:
[LIST OFFICIAL SITES, FILINGS, product docs, credible publications, uploaded files, or connected sources]

Instructions:
1. Map the market into segments, buyer personas, vendor types, and adoption drivers.
2. Distinguish Evidence, Inference, Uncertainty, and Unsupported claims.
3. Preserve citations and source links for all factual claims.
4. Do not invent market sizes, rankings, customer counts, pricing, or product capabilities.
5. Identify missing evidence and items requiring human or expert review.
6. Return: segment map, vendor categories, demand signals, risks, unanswered questions, and recommended next validation steps.

Required inputs

• The market definition and what is excluded.
• The decision the landscape should support.
• Allowed domains, uploaded files, or connected-source boundaries.

Expected output

You should receive a structured market map that separates documented evidence from hypotheses. The report should identify where customer interviews, financial diligence, legal review, or technical validation are needed before a go-to-market or investment decision.

Verification checkpoint

Check that the market definition did not expand mid-run. If the report includes numbers, rankings, or vendor capability claims, verify the citation and remove any figure that is not directly supported.

Prompt 15: Vendor Scorecard

Purpose

Use this to compare vendors against explicit criteria while avoiding the common error of treating marketing pages as complete evidence. The prompt is useful for procurement, security review, enterprise administration, and technical shortlisting.

Copy-paste prompt

Build a source-controlled vendor scorecard.

Vendors to compare:
[LIST VENDORS]

Use case:
[DESCRIBE BUSINESS OR TECHNICAL USE CASE]

Evaluation criteria:
[SECURITY, ADMIN CONTROLS, INTEGRATIONS, DATA HANDLING, SUPPORT, ACCESSIBILITY, COST MODEL, DEPLOYMENT FIT, etc.]

Evidence rules:
[STATE ALLOWED SOURCES AND EXCLUDED SOURCES]

Instructions:
1. Evaluate only against the listed use case and criteria.
2. Distinguish Evidence, Inference, Uncertainty, and Unsupported claims.
3. Preserve citations and source links for every scored item.
4. If a criterion is not documented, mark it as “not found in reviewed sources” instead of guessing.
5. Identify conflicts between sources and any items requiring human or expert review.
6. Return a scorecard with criteria, evidence notes, citation links, confidence, risks, and follow-up questions for each vendor.

Required inputs

• The vendor list.
• The use case and non-negotiable requirements.
• Allowed sources, such as official docs, security pages, procurement documents, or uploaded RFP responses.

Expected output

The scorecard should expose both strengths and evidence gaps. A vendor should not receive a high score for a criterion unless the reviewed sources actually document the capability or policy relevant to your use case.

Verification checkpoint

Before sharing with procurement or leadership, confirm that every disqualifying concern and every high score has a source. Treat pricing, retention, compliance, and permission claims as requiring direct source verification or vendor confirmation.

Prompt 16: Policy Comparison

Purpose

Use this to compare policies, terms, standards, internal procedures, or regulatory proposals without converting nuanced text into false equivalence. It works best when you provide the exact documents or source boundaries.

Copy-paste prompt

Compare the following policies using a source-controlled method.

Policies or documents:
[LIST OR ATTACH DOCUMENTS]

Comparison objective:
[STATE WHY THE COMPARISON IS NEEDED]

Scope limits:
[JURISDICTION, ORGANIZATION, DATE RANGE, BUSINESS UNIT, PRODUCT, RISK AREA]

Instructions:
1. Compare definitions, obligations, permissions, prohibitions, exceptions, enforcement mechanisms, and review requirements.
2. Distinguish Evidence, Inference, Uncertainty, and Unsupported claims.
3. Preserve citations, source links, section references, or document locations.
4. Do not provide legal conclusions; identify legal or compliance questions requiring qualified review.
5. Highlight contradictions, ambiguity, missing definitions, and operational impact.
6. Return a comparison table, decision implications, unresolved questions, and expert-review checklist.

Required inputs

• The policies, terms, standards, or official sources to compare.
• The jurisdiction or organizational boundary.
• The operational question, such as launch readiness, vendor approval, or internal policy update.

Expected output

The answer should distinguish textual differences from practical differences. For example, two policies may use different language but impose similar approval steps, while two similar-looking clauses may differ materially because one includes an exception or review trigger.

Verification checkpoint

Have counsel, compliance, or the accountable policy owner review any interpretation that affects obligations, risk acceptance, customer commitments, employee monitoring, regulated data, or external publication.

Prompt 17: Customer Pain-Point Synthesis

Purpose

Use this to synthesize interviews, support tickets, sales notes, community posts, or survey excerpts into a defensible customer-pain map. The prompt prevents the model from overstating frequency or turning anecdotes into market-wide conclusions.

Copy-paste prompt

Synthesize customer pain points from the provided source set.

Source set:
[PASTE OR ATTACH INTERVIEWS, TICKETS, NOTES, SURVEY EXCERPTS, CALL SUMMARIES, OR APPROVED PUBLIC SOURCES]

Customer segment:
[DEFINE SEGMENT]

Decision to support:
[ROADMAP, POSITIONING, SALES ENABLEMENT, CHURN REDUCTION, SUPPORT PRIORITIZATION, etc.]

Instructions:
1. Identify recurring pain points, jobs-to-be-done, workarounds, objections, and desired outcomes.
2. Distinguish Evidence, Inference, Uncertainty, and Unsupported claims.
3. Preserve citations, source links, ticket IDs, interview labels, or document references.
4. Do not infer prevalence beyond the reviewed source set unless supported by quantified evidence.
5. Identify sensitive information, confidentiality concerns, and items requiring human or expert review.
6. Return themes, representative evidence, confidence, segment differences, product implications, and validation questions.

Required inputs

• The customer source set and any confidentiality boundary.
• The customer segment and decision context.
• Rules for quoting, anonymizing, or excluding sensitive material.

Expected output

The synthesis should include theme clusters with representative evidence, not just a list of complaints. A useful result separates observed pain from inferred root cause and marks whether a theme appears in one source, several sources, or a quantified dataset.

Verification checkpoint

Before reuse in a roadmap, pitch deck, or customer-facing artifact, remove sensitive information, confirm rights and permissions for included material, and retain source references without exposing confidential details.

Prompt 18: Implementation-Playbook Conversion

Purpose

Use this to turn a research report into a practical rollout plan, 30/60/90-day roadmap, risk register, or operating checklist. The prompt forces the model to separate proven findings from assumptions that require owner review.

Copy-paste prompt

Convert the research findings into an implementation playbook.

Research findings:
[PASTE REPORT OR LINKED FINDINGS AVAILABLE IN THE CURRENT TASK]

Implementation goal:
[STATE GOAL]

Organization context:
[TEAM SIZE, SYSTEMS, CONSTRAINTS, REGIONS, STAKEHOLDERS, APPROVAL REQUIREMENTS]

Output format:
[30/60/90-DAY PLAN, RISK REGISTER, RACI, CHECKLIST, OPERATING MODEL, CHANGE PLAN, etc.]

Instructions:
1. Translate findings into phases, actions, owners, dependencies, risks, and decision gates.
2. Distinguish Evidence, Inference, Uncertainty, and Unsupported claims.
3. Preserve citations and source links for findings that drive each action.
4. Mark assumptions that require human validation before execution.
5. Identify items requiring expert review, including security, legal, finance, compliance, data governance, medical, or safety implications.
6. Do not invent approvals, permissions, budgets, dates, staffing, product capabilities, or legal conclusions.
7. Return an implementation playbook with evidence-linked actions, open questions, review owners, and go/no-go checkpoints.

Required inputs

• The research report or summarized findings.
• The implementation goal and format.
• Organizational constraints, approval needs, and known dependencies.

Expected output

The output should be executable but not autonomous. It should tell a human team what to review, approve, sequence, and validate before acting, especially where the research artifact depends on citations, internal files, connected-source material, or unresolved assumptions.

Verification checkpoint

Open any generated document, spreadsheet, or link and confirm it exists, is readable, and preserves the required citations. OpenAI’s Deep Research guidance emphasizes verification of underlying sources and final artifacts, especially before decisions, external sharing, or high-stakes use.

Prompts 19–25: Turn Research Into Decision Artifacts, Then Gate It for Safe Use

Prompts 19–25 convert Deep Research output into executive, presentation, spreadsheet, compliance, citation, and publication-ready artifacts. OpenAI’s Deep Research guidance emphasizes defining the desired outcome, steering the run, preserving citations, and verifying underlying sources before decisions or external sharing; these templates turn those requirements into reusable operating controls rather than one-off cleanup steps.

The article provides a 2026 decision framework for choosing among AI coding assistants such as Copilot, Cursor, Claude Code, Codex, and Windsurf. The complete How to Choose an AI Coding Assistant in 2026: Decision Framework for Copilot, Cursor, Claude Code, Codex, and Windsurf article provides the destination-specific detail for this section’s Decision Memo Prompt Guide decision because although domain-specific to coding assistants, it aligns with decision-memo workflows by showing how to structure comparative analysis into a practical decision framework.

Operational rule: treat every final artifact as a draft until a human has opened the cited sources, checked sensitive information, confirmed rights and permissions for quoted or copied material, and verified that the generated file or link exists and opens successfully when an editable deliverable is requested.

Prompt 19: Executive Decision Memo

Purpose: Use this prompt when a research run must become a concise decision memo for an executive, board, investment committee, product council, security steering group, or procurement committee. The memo should separate the recommendation from the evidence and make clear where judgment, missing information, or expert review remains.

Copy-paste prompt:

Convert the completed Deep Research findings into an executive decision memo.

Decision to support: [insert decision]
Audience: [insert audience]
Decision deadline: [insert date or timeframe]
Options to compare: [insert options]
Risk tolerance: [insert risk posture]
Required sections:
1. One-paragraph recommendation
2. Decision context
3. Options considered
4. Evidence summary with citations preserved
5. Key risks and mitigations
6. Open questions
7. Items requiring human or expert review
8. Final decision log entry

Distinguish evidence, inference, uncertainty, and unsupported claims. Preserve citations next to the claims they support. Identify any statements that require human, legal, financial, technical, security, medical, or domain-expert review before the memo is used for a real decision. Do not make the final decision for us; prepare the decision artifact.

Required inputs: Provide the decision question, audience, options, risk posture, deadline, and any non-negotiable constraints such as budget, jurisdiction, security baseline, policy deadline, customer segment, or integration requirement.

Expected output: A memo with a clear recommendation, evidence-backed rationale, stated tradeoffs, cited claims, unresolved questions, and a decision-log section that can be copied into a governance system or meeting packet.

Verification checkpoint: Confirm that each recommendation rests on cited evidence or explicitly labeled judgment. If the memo contains legal, financial, medical, safety, security, or regulated-market conclusions, route it to the appropriate human reviewer before relying on it.

Prompt 20: Presentation Brief

Purpose: Use this prompt to convert research into a slide-ready brief without asking the model to invent charts, metrics, or visuals that were not supported by the source material. This is useful for leadership readouts, customer discovery summaries, market scans, policy comparisons, and vendor evaluations.

Copy-paste prompt:

Create a presentation brief from the Deep Research findings.

Presentation audience: [insert audience]
Meeting objective: [inform / decide / align / approve / challenge]
Time available: [insert minutes]
Preferred slide count: [insert count or range]
Required slide outline:
1. Title and decision question
2. Executive takeaway
3. What we investigated
4. Evidence-backed findings
5. Contradictions or weak signals
6. Options or implications
7. Risks, dependencies, and mitigations
8. Recommendation or next step
9. Appendix: source list and citation notes

For each slide, provide the slide title, three to five bullet points, suggested visual only if supported by the evidence, and speaker notes. Distinguish evidence, inference, uncertainty, and unsupported claims. Preserve citations in speaker notes and appendix entries. Identify claims, visuals, or recommendations that require human or expert review before presentation.

Required inputs: Supply the audience, meeting purpose, timebox, desired slide count, and any brand, template, confidentiality, or regional constraints that affect what can be shown.

Expected output: A slide-by-slide brief with concise bullets, speaker notes, citation references, appendix material, and warnings where the research does not support a proposed chart or claim.

Verification checkpoint: Check that no slide converts a tentative finding into a fact. If a chart is suggested, verify that the underlying figures appear in the cited sources and that the chart does not combine incompatible datasets.

Prompt 21: Spreadsheet Evidence Register

Purpose: Use this prompt when the team needs a structured evidence register for audit, procurement, due diligence, research operations, or quality review. A spreadsheet format helps reviewers filter claims by source type, confidence, owner, follow-up task, and publication readiness.

Copy-paste prompt:

Convert the research findings into a spreadsheet-ready evidence register.

Use the following columns:
- Claim ID
- Claim
- Evidence type
- Citation or source link
- Source title or document name
- Source date if available
- Source owner or publisher if available
- Evidence strength: high / medium / low
- What the evidence directly supports
- Inference made from the evidence
- Uncertainty or limitation
- Unsupported claim flag
- Human or expert review required
- Reviewer owner
- Follow-up action
- Publication status: keep / revise / remove

Return the register as a table that can be copied into a spreadsheet. Distinguish evidence, inference, uncertainty, and unsupported claims. Preserve citations in the citation column and do not merge multiple sources into one citation unless the claim is genuinely supported by all of them. Identify all rows requiring human or expert review.

Required inputs: Provide the research report, preferred taxonomy for evidence strength, reviewer names or roles, and whether the register is for internal review, external publication, investment diligence, compliance, or procurement.

Expected output: A spreadsheet-ready table mapping each important claim to a source, evidentiary limitation, owner, follow-up action, and publication status.

Verification checkpoint: Sort the register by “unsupported claim flag” and “human or expert review required” before reuse. Remove or revise claims that cannot be tied to a source or a clearly labeled inference.

Prompt 22: Sensitive and Confidential-Information Review

Purpose: Use this prompt before sharing research outside the original workspace, team, or intended audience. OpenAI’s guidance for external sharing includes rechecking citations and removing sensitive information; this prompt turns that advice into a structured review pass.

Copy-paste prompt:

Review the draft artifact for sensitive, confidential, restricted, or audience-inappropriate information.

Artifact type: [memo / slides / spreadsheet / report / email / policy brief]
Intended audience: [internal team / executives / customer / investor / public / regulator / vendor]
Information to flag:
- Personal data
- Customer or employee details
- Non-public financial or commercial information
- Security-sensitive architecture, incidents, credentials, or vulnerabilities
- Confidential product plans
- Contract terms or pricing not cleared for sharing
- Internal deliberations
- Licensed or third-party material with unclear permissions
- Any workspace, connected-app, or uploaded-file content that may not be approved for redistribution

For each flagged item, provide the exact passage, sensitivity category, reason for concern, recommended action, and reviewer role. Distinguish evidence, inference, uncertainty, and unsupported claims. Preserve citations for retained claims. Identify items requiring human, legal, security, privacy, procurement, communications, or executive review before sharing.

Required inputs: Provide the draft artifact, intended audience, allowed disclosure level, and any company confidentiality classifications or policy terms the review should apply.

Expected output: A redline-style risk table identifying passages to keep, revise, remove, anonymize, aggregate, or escalate to a named review function.

Verification checkpoint: Do not rely on the model as the final privacy or confidentiality authority. A designated human owner should inspect the artifact and confirm that workspace, contract, and data-handling obligations are met before distribution.

Prompt 23: Copyright and Quotation Review

Purpose: Use this prompt when a report contains quotations, copied tables, screenshots, long excerpts, third-party frameworks, or source-heavy language. The goal is to reduce rights and attribution risk, not to obtain a legal opinion from the model.

Copy-paste prompt:

Review this draft for copyright, quotation, attribution, and permissions issues before external or broad internal sharing.

Artifact: [paste artifact or attach draft]
Sharing context: [internal / customer-facing / public / investor / training / sales / policy]
Flag the following:
- Long copied passages
- Direct quotations without clear attribution
- Paraphrases that are too close to source wording
- Tables, images, screenshots, diagrams, or frameworks that may need permission
- Missing citations
- Claims where citation exists but does not appear to support the wording
- Material that should be summarized instead of quoted

For each issue, provide the passage, source if known, issue type, risk rationale, suggested rewrite or removal, and whether human or legal review is required. Distinguish evidence, inference, uncertainty, and unsupported claims. Preserve citations for retained material and identify all items requiring expert rights, legal, communications, or editorial review.

Required inputs: Provide the artifact, sharing context, citation list, and any known rights status for third-party materials such as licensed analyst reports, vendor documentation, standards, survey data, or images.

Expected output: A rights-review table with recommended rewrites, removals, attribution fixes, and legal-review flags for material that should not be published without permission analysis.

Verification checkpoint: Treat the result as an editorial triage, not legal clearance. Confirm permissions, license terms, quotation policy, and attribution requirements with qualified reviewers before external publication.

Prompt 24: Citation Spot-Check

Purpose: Use this prompt after a report, memo, or slide deck is drafted and before it is trusted. OpenAI states that citations improve traceability but do not guarantee correctness, so a citation audit must test whether each cited source actually supports the claim.

Copy-paste prompt:

Perform a citation spot-check on the draft artifact.

Artifact: [paste artifact]
Spot-check scope: [all citations / highest-risk claims / random sample of N / claims used in recommendation]
For each checked claim, evaluate:
1. The exact claim made
2. The citation attached to it
3. Whether the cited source directly supports the claim
4. Whether the claim overstates, understates, or changes the source
5. Whether the claim needs additional sources
6. Whether the source is primary, secondary, vendor-authored, user-provided, or otherwise limited
7. Recommended fix

Use labels: supported, partially supported, unsupported, citation missing, citation mismatch, expert review required. Distinguish evidence, inference, uncertainty, and unsupported claims. Preserve citations that pass review and identify all claims requiring human or expert verification before decisions or publication.

Required inputs: Provide the artifact, source list, desired sampling method, and the risk categories that matter most, such as financial projections, safety statements, legal interpretation, medical claims, security assertions, or product capability statements.

Expected output: A citation-audit table showing which claims are supported, partially supported, unsupported, mismatched, or in need of expert review.

Verification checkpoint: Human reviewers should open the underlying sources for every high-impact claim. Remove claims that cannot be verified, and downgrade strong language when the source only supports a narrower statement.

Prompt 25: Final Publication Gate

Purpose: Use this prompt as the final gate before sending, publishing, presenting, filing, or handing a research artifact to another team. It combines evidence quality, citation support, confidentiality, rights review, artifact usability, and human approval into one release checklist.

Copy-paste prompt:

Run a final publication gate on this research artifact.

Artifact type: [report / memo / slides / spreadsheet / brief / email / other]
Destination: [internal / executive / customer / vendor / regulator / public]
Required gate checks:
- Decision question is clear
- Audience and scope are stated
- Evidence is separated from inference
- Uncertainty is visible
- Unsupported claims are removed or labeled
- Citations are preserved and attached to supported claims
- Sensitive or confidential information is removed or approved
- Copyright, quotation, and permissions concerns are resolved or escalated
- High-impact claims have named human reviewers
- Generated files, links, tables, or spreadsheets have been opened and inspected
- Remaining risks and open questions are listed
- Final owner and approval status are explicit

Return a release checklist with status: pass / revise / block / expert review required. Distinguish evidence, inference, uncertainty, and unsupported claims. Preserve citations in the final version. Identify all items requiring human, expert, legal, security, privacy, editorial, financial, medical, or executive review before release.

Required inputs: Provide the final artifact, destination audience, approval owners, risk categories, and any publication rules such as embargoes, customer confidentiality, workspace retention obligations, or brand-review requirements.

Expected output: A gate table with pass, revise, block, or expert-review status for each release criterion, plus a short final owner checklist.

Verification checkpoint: Do not publish solely because the model returns “pass.” A human owner should confirm source support, artifact integrity, permissions, confidentiality, and approval status before release.

Implementation Guidance for Source-Controlled Prompt Operations

Version these prompts the same way you version policy templates, research protocols, and release checklists. Store the prompt text, date, owner, intended use case, allowed source classes, required artifact format, and verification steps so reviewers can tell which version produced a memo, slide brief, spreadsheet register, or publication gate.

Control	Practical implementation	Failure prevented
Prompt version	Use a visible version string such as “DR-GATE-2026-09-10-v1” in the prompt header and artifact footer.	Prevents teams from mixing old review standards with current publication rules.
Source boundary	State allowed domains, uploaded files, connected sources, and excluded source types before the run starts.	Prevents broad web material from being treated as if it came from approved sources.
Allowance awareness	Before a long run, decide whether the task belongs in Chat, Work, or Codex and confirm that the workspace, plan, credits, tools, and connected-source permissions support the request.	Prevents incomplete research caused by unavailable tools, exhausted allowance, or missing permissions.
Human review	Assign named reviewers for legal, security, privacy, financial, medical, technical, editorial, or executive approval where relevant.	Prevents citations and polished wording from being mistaken for final authority.
Recovery path	When output fails, rerun from the last verified checkpoint rather than asking for a broad rewrite.	Prevents the model from compounding earlier citation, scope, or sensitivity errors.

Allowance awareness matters because OpenAI documents that Deep Research uses plan-dependent task allowance in Chat, while Work usage consumes the existing Work/Codex allowance or credits rather than the separate Chat Deep Research task allowance. Treat long market scans, multi-source diligence, and artifact generation as budgeted tasks, especially when a workspace has role-based access controls, web-search requirements, connected-app permissions, or credit governance.

Human review should be assigned before the research run begins, not after a polished artifact appears. A research lead can approve evidence scope, a security reviewer can inspect confidential or sensitive content, counsel can assess rights and regulated claims, and an executive owner can decide whether the recommendation is acceptable under the organization’s risk tolerance.

Failure recovery should start with classification. If citations are missing, run Prompt 24 before rewriting. If confidential information appears, run Prompt 22 and remove or mask material before producing slides. If the artifact is too broad, return to the decision question and source boundary rather than asking for a shorter summary. If the generated spreadsheet, document, presentation, or link does not open, request a regenerated artifact only after preserving the text output and evidence register.

Do not describe ChatGPT-5.5 as the Deep Research model for every account, surface, plan, or workspace. Use these templates as the control layer for framing, steering, evaluating, and packaging research while preserving OpenAI’s documented boundaries: source access depends on available tools and permissions, citations require verification, output formats depend on supported capabilities, and final publication remains a human responsibility.

Useful Links

• Deep Research in ChatGPT
• OpenAI Academy Deep Research resource
• ChatGPT Work and Codex
• GPT-5.5 model documentation