7 Stock Research Prompts for Claude, Cursor, Codex, and AI Agents

A broad prompt produces a broad answer. “Analyze this stock” gives an AI agent too much freedom to choose the time horizon, evidence, comparison set, and definition of risk. The result may sound polished while hiding the parts you actually needed to decide.

Better stock-research prompts specify the job, the evidence to retrieve, and the shape of the answer. When the agent can use a read-only stock-market MCP server, it can retrieve current structured context instead of answering only from memory.

Below are five reusable prompts for Claude Code, Cursor, Codex, or another tool-connected agent. They are designed around Stocksbrew for Agents, but the prompt patterns work with any trustworthy research data source.

Before you prompt: set three boundaries

Tell the agent the time horizon, the decision you are preparing, and what it must not assume. A long-term investor reviewing a holding has a different question from someone comparing two candidates before earnings.

A useful preamble is:

Use current tool data and show the as-of time. Separate facts from interpretation. Call out missing or stale information. Do not give a trade instruction; give me the evidence, risks, and conditions that would change the conclusion.

That one paragraph improves nearly every research prompt below.

1. The structured single-stock review

Review MSFT for a 12-to-24-month holding period. Use the current stock-intelligence tool. Organize the answer into: current call, price and target room, business drivers, valuation, quality, growth, momentum, catalysts, risks, and invalidation conditions. Show the as-of time and flag any weak or missing evidence.

This prompt prevents the agent from collapsing every dimension into a single bullish or bearish label. It also makes the invalidation conditions explicit: what would have to change for the current view to stop making sense?

2. The apples-to-apples comparison

Compare NVDA, AMD, and AVGO using the same fields for every company: valuation, growth, quality, momentum, one-year return, target upside, technical context, and the strongest risk. Rank the evidence quality separately from the investment attractiveness. Do not treat a missing field as zero.

The last sentence is important. Missing data is not bad data; it is unknown data. The agent should preserve that distinction. Use a short comparison list because a five-stock table is easier to audit than a fifty-stock ranking.

3. The bull case, bear case, and decision hinge

For TSLA, build the strongest evidence-backed bull case and bear case from the current Stocksbrew context. Then identify the three decision hinges that would move a neutral investor toward either side. Distinguish company fundamentals from market sentiment and price momentum.

This prompt is useful when a stock attracts strong narratives. Asking for the strongest version of both sides reduces the chance that the agent simply mirrors the tone of your question.

4. The freshness and evidence audit

Review the current Stocksbrew record for META. Show the as-of time and freshness for the response. List any missing, aging, or weak fields. Then tell me which parts of the conclusion depend on those fields and what primary source I should verify next.

This prompt checks the quality of the context before debating the conclusion. Missing evidence should stay missing; the agent should not smooth over it with confident prose.

5. The evidence-gap follow-up

Review AAPL using current Stocksbrew stock intelligence. After the first answer, identify the three claims supported by the strongest evidence and the three claims supported by the weakest evidence. Do not fill any gap from memory without labeling it as outside the Stocksbrew response.

This keeps tool output separate from model interpretation. It is especially useful when a polished summary makes every claim sound equally certain.

Bonus: the pre-earnings risk checklist

Review AMZN ahead of its next earnings event. Using current stock intelligence, list the key operating drivers, valuation sensitivity, major risks, and the evidence to check immediately after the report. Do not predict the exact price move.

The goal is preparation, not a point forecast. A checklist lets you compare the reported result with the assumptions that mattered before the event. For a broader manual workflow, use the earnings report analysis guide.

Follow-up prompts that improve the answer

The first answer should start a research process, not end it. Useful follow-ups include:

• “Which conclusion depends on the weakest data?”
• “What evidence would make you reverse this ranking?”
• “Separate facts returned by the tool from your interpretation.”
• “Which number is most sensitive to a stale timestamp?”
• “What primary source should I verify next?”
• “Rewrite this as a checklist I can use after earnings.”

These follow-ups make uncertainty visible. They are more useful than asking the model to become “more confident.”

Prompts to avoid

Avoid prompts such as “Tell me the best stock,” “What will double next?” or “Should I buy this today?” They hide your time horizon and risk tolerance, encourage false precision, and invite the agent to convert incomplete research into a directive.

Also avoid asking for hundreds of tickers at once. Bounded comparison sets are easier to audit and less likely to turn a personal research tool into an accidental bulk-data workflow.

How to give the prompts current context

A prompt alone does not make an answer current. The agent needs a data connection. Stocksbrew Pro members can add the hosted MCP URL to a compatible client or create a personal token for the REST API. The core stock tools cover search, one-stock intelligence, and comparisons.

Every successful response includes a schema version, as-of time, freshness information, an informational-use notice, and current usage. The agent cannot trade or change your account through these tools.