The AI research plan review
AI research agents can gather sources, compare evidence, and draft reports. The useful skill is reviewing the research plan before the agent starts collecting evidence.
AI research tools are getting more autonomous. Google describes Deep Research agents that can work across the web and custom sources, generate charts, and support collaborative planning before execution. OpenAI's research index shows frontier models being used for serious scientific and mathematical work, while its research and analysis hub tracks how people and organizations are adopting AI in practice.
For everyday knowledge workers, the lesson is straightforward: do not start with "research this." Start by making the AI show the research plan. The plan is where you catch bad scope, missing sources, weak definitions, and output that will not help the decision.
The skill
A research plan review is a short checkpoint before the AI begins a deeper investigation. It defines the question, source types, exclusions, evidence standards, output format, and decision the research needs to support.
AI research plan review
Research question:
{the decision or question this research supports}
Audience:
{who will use the answer}
Decision needed:
{what someone will decide after reading the research}
Source plan:
- Primary sources:
- Expert or market sources:
- Internal sources:
- Sources to avoid:
Evidence standard:
{what counts as strong evidence, weak evidence, or unsupported}
Output format:
{brief, table, memo, slide outline, recommendation, risk list}
Known assumptions:
- {assumption to test}
- {assumption to test}
Stop condition:
{when the AI should report that evidence is insufficient}
A worked example: vendor comparison
A weak request is: "Research the best AI meeting tool for our team." That invites the model to browse broadly, over-weight marketing pages, and produce a generic ranking.
The plan-review version is more useful:
Research question:
Which AI meeting tool should a 25-person operations team trial for decision logs and follow-up tracking?
Audience:
Operations lead and finance approver.
Decision needed:
Choose one tool for a 30-day pilot, or decide not to pilot yet.
Source plan:
- Primary sources: vendor pricing pages, product docs, security pages
- User evidence: recent reviews from teams using meeting transcripts
- Internal sources: current meeting workflow, privacy policy, must-have feature list
- Avoid: affiliate roundups, outdated review posts, unsourced comparison tables
Evidence standard:
- Strong: current vendor docs or direct user evidence with date
- Weak: generic blog claims or missing dates
- Unsupported: feature claims with no link
Output format:
1. Short recommendation
2. Comparison table
3. Risks and gaps
4. Pilot checklist
Stop condition:
If pricing, retention, or consent controls are unclear, flag the gap and do not recommend a purchase.
The prompt
Use this before asking an AI research tool to browse, analyze, or write a report:
Before doing the research, create a research plan for my review.
Question:
{what I need to understand}
Decision this supports:
{what will be decided after the research}
Audience:
{who will read it}
Create a plan with:
1. Scope
2. Source types to use
3. Sources to avoid
4. Search terms or sub-questions
5. Evidence standards
6. Output format
7. Known assumptions to test
8. Stop condition if evidence is weak
Do not begin the research yet. Wait for me to approve or edit the plan.
The review checklist
- Question: Is the research tied to a real decision, or just curiosity?
- Scope: Is the topic narrow enough for useful evidence?
- Sources: Are primary sources included and weak sources excluded?
- Assumptions: Does the plan test what you already believe?
- Output: Will the final format help someone act?
- Stop condition: Will the AI admit when evidence is too thin?
Why it works
Research agents can do more work than a human would casually ask for. That is the gift and the danger. If the plan is wrong, the agent may produce a polished report from the wrong evidence.
The plan review moves your judgement earlier in the workflow. You guide the search path before the AI spends time gathering sources, and you make the final report easier to verify because the evidence rules were set upfront.