AI ETF Research: A Hands-On Workflow with Copy-Ready Prompts
A source list, a fact ledger, a growth-or-income question map, and a short research memo with unresolved issues. You will not make an automated buy list. The basic workflow needs no additional skill or MCP connection.
Turn official ETF documents into a research memo that follows your investment objective. These copy-ready prompts show what to ask, what a useful answer should contain, and what you must still check yourself.
Image above: AI-generated editorial illustration of document analysis and evidence checking—not an actual product benchmark or investment result.
The ETF category explains how a fund works and how to judge its role. AI Tools takes the next step: opening the right documents, asking better questions, and producing research you can inspect. A fluent answer is not the destination. A usable, traceable report is.
Our earlier guide to delegating and verifying ETF research develops the principles. This article is the companion workshop. We use SCHD as a concrete source-finding example, not as a recommendation. You can replace it with another U.S.-listed ETF once you identify that fund’s official documents.
1. Choose the environment and the research job
A model and a workspace are different things. GPT-6.1 Sol is a model; ChatGPT Work and Codex are environments in which you can carry out tasks. Claude is another service whose controls and available models may differ. Do not confuse the name of a model with permission to browse, read files, or save results.
Main route: ChatGPT Work
Start a new task, select GPT-6.1 Sol if it is available in your account, and attach the public documents you want examined. Paste the prompts below one at a time, reviewing each result before continuing.
Use it for: a source-backed memo and files you can review. No API key or custom connection is needed for the basic exercise.
Optional route: Codex
Use a dedicated research folder that contains only public material. Ask Codex to preserve the originals and save the ledger and memo as separate files. Check which local-file and browsing tools your environment actually offers.
Use it for: repeatable file organization. Coding or brokerage access is not part of this exercise.
In Claude, begin with a new conversation and attach the same files if your account supports them. The plain-language prompts are portable starting points, not a claim that the services have identical features or will produce identical answers. This article does not provide a tested Claude-model ranking or call an unverified model “the latest.”
Before running anything, record the service, visible model name, date, search setting, and filenames. If you cannot find a control, record that limitation rather than guessing. Product availability and interfaces change; the instructions here are task-based, not a promise that every account has the same buttons.
The official-source facts in the worked example below were checked during editorial research on October 10, 2026. The report excerpt is an editorially prepared demonstration, not a captured response from a separate GPT or Claude test. These prompts are proposed research instructions, not a measured performance claim. Your output will vary.
2. Find official sources before asking for a conclusion
Open the official SCHD product page, then its document library. For the first pass, save the product-page information and a current fact sheet. Add the prospectus for strategy and risk questions. For holdings overlap or income research, add the relevant dated holdings or distribution records. Save publication dates separately from the day you downloaded a file.
You do not need every available document to start. You do need the documents that can answer your question. A fact sheet is a useful summary, but it is not a substitute for reading the relevant prospectus sections when a strategy or risk claim matters.
If search is available in your tool, use Prompt 1 to organize discovery. If it is not, open the issuer’s website yourself and attach the documents. A pasted link does not guarantee that an assistant can read the page. If it reports an access failure, supply a saved public PDF or a clearly labeled excerpt; do not accept an answer assembled from memory.
Help me build a public-source research packet for SCHD, the Schwab U.S. Dividend Equity ETF. Use the fund issuer's official website first. Find the product page, current fact sheet, prospectus, dated holdings, and distribution records relevant to this task. For each item, give its title, official URL, publication or data date, access date, and the question it can answer. Distinguish publication dates from download dates. Open the source if your tools permit. Mark anything you cannot access as NOT READ. Do not invent a document, page number, or fact. Do not recommend buying, selling, or allocating money. Stop after the source list so I can review it.
Expected result: a source list you can open yourself, with inaccessible items clearly identified. An impressive list of links is not enough; inspect the issuer, document title, and date. You can copy the prompt manually if your browser blocks the copy button.
3. Extract facts into an evidence ledger
Once the packet is attached, separate extraction from interpretation. Ask the assistant to read the files it actually has, not to create a broad market outlook. Each important claim needs a place where a person can check it.
Use only the public source packet attached to this conversation. First list the files or excerpts you can and cannot read. Build a fact ledger for SCHD. Include the investment objective, tracked index, stated expense ratio, strategy, main disclosed risks, and any holdings or distribution facts relevant to my question. For every material claim, give: fact; value or wording; date and unit if applicable; document title; page, section, or table; and verification status. Use FOUND IN SOURCE, NOT FOUND, or CONFLICTING SOURCES. Do not call a fact independently verified merely because you extracted it. Preserve the exact meaning of yield labels. Do not substitute distribution yield for total return or project future payouts. Separate your interpretation from documented facts. If evidence is unavailable, say so. Stop before writing a recommendation.
A small answer key you can check
Official SCHD product page, checked October 10, 2026. This is an editorial extraction example, not a screenshot of an AI response. Live page fields may change.
- Objective / index
- Tracks the Dow Jones U.S. Dividend 100 Index’s total return before fees and expenses. Check the page’s Objective and Fund Profile sections.
- Expense ratio
- 0.060%, shown in Fund Profile. That is the stated fund expense ratio—not a promise of your total trading cost.
- Style classification
- Fund Profile lists the Morningstar category as Large Value. Do not relabel it a growth-stock strategy simply because an investor seeks capital growth.
- Unresolved question
- Whether next year’s distributions will cover a particular household’s expenses. The source fields above do not establish that outcome.
Source: Schwab Asset Management. No issuer chart or image is reproduced.
This is deliberately a small checkpoint. A correct expense ratio does not prove that the rest of a memo is sound. Open the cited section for every claim that could materially change your decision. If two sources disagree, preserve their dates and investigate the discrepancy rather than averaging the numbers.
4. Ask different questions for growth and income
The same fund can be studied for different portfolio roles. That does not make all strategies interchangeable. Capital growth is an investor objective; growth-stock exposure is a style. Dividend growth, high dividend yield, and option-related distributions also require different questions.
Capital-growth objective
Ask: What market exposure, concentration, and overlap am I taking on? What could challenge the holding period or the fund’s intended role?
Do not infer: a long horizon means unlimited capacity to absorb losses.
Income objective
Ask: What generates distributions? How variable might they be? What additional records are needed to examine them alongside total return and capital risk?
Do not infer: a dividend-oriented strategy guarantees a household’s future cash flow.
Personal tendencies belong in the report as behavior to plan for, not as a shortcut to a ticker. “I check prices every day” can lead to a written review schedule. It does not establish that a dividend fund is safer or more suitable. State financial capacity separately from emotional tolerance.
Use the source ledger above to build research questions, not a suitability judgment. My hypothetical research profile: - Objective: [capital growth / dividend growth / current income / mixed; replace this]. - Horizon and first expected cash need: [replace]. - Distributions: [reinvest / spend / undecided]. - Financial capacity for loss: [describe broadly, or write NOT ASSESSED]. - Behavior to plan for: [for example, frequent price checking or impulsive switching]. - Existing portfolio exposure: [broad categories only, or UNKNOWN]. Do not request account numbers, tax records, or private statements. Create separate capital-growth and income question lists. Explain which list fits my stated research objective and which overlaps remain. Distinguish growth objectives from growth-stock styles and dividend growth from high yield or option income. Identify additional documents required for unanswered questions. Do not infer risk capacity from personality or horizon alone. Do not recommend a ticker, trade, or allocation.
For your first practice run, use a hypothetical profile: a ten-year horizon, distributions reinvested, no near-term spending planned, existing holdings not supplied, and loss capacity not assessed. The appropriate output should identify the missing information—not pretend the profile is complete. For the underlying investment reasoning, see our growth-versus-dividend management guide.
5. Build and verify your research memo
Now ask for a report. “Personalized” should mean the questions follow your stated objective and constraints. It should not mean an assistant has established that an investment is suitable. The report is research support; a decision remains yours.
Using the attached packet, fact ledger, and research profile, write a 700–1,000-word ETF research memo in plain American English. Use these headings: 1. Research objective and missing investor information 2. Documented fund facts, with dates and source locations 3. Questions for capital growth versus income 4. Main risks, portfolio-overlap questions, and evidence gaps 5. What a human must verify before making a decision Separate SOURCE FACT, INTERPRETATION, and UNRESOLVED wherever the distinction matters. Include a compact source list. Preserve uncertainty. Do not forecast returns or dividends, claim suitability, or give buy/sell/weight instructions. If required information is unavailable, deliver a limited memo and explain what is missing. Return the memo for review. Do not publish, email, connect to a broker, or modify source files.
What a useful passage looks like
Editorially prepared demonstration using the source checkpoint and hypothetical profile above—not a captured model output.
SOURCE FACT: The issuer identifies a dividend-oriented index and a Large Value category. These labels describe the fund, not the investor’s required outcome.
INTERPRETATION: For a capital-growth objective with distributions reinvested, the research should examine total-return exposure, concentration, and existing holdings—not simply the size of a cash payout.
UNRESOLVED: Existing holdings and loss capacity were not supplied. This memo cannot establish portfolio overlap, financial resilience, or suitability.
A narrower report with visible gaps is more useful than a persuasive conclusion supported by invisible assumptions. Before you accept it, work through this five-step check.
- 1
Open the cited source
Confirm that it exists, is official where claimed, and was actually read.
- 2
Match fact, date, and unit
Check each consequential figure and its label. Preserve different data dates.
- 3
Separate evidence from inference
Ask whether a conclusion follows from the source or adds a forecast the source cannot support.
- 4
Check purpose and omissions
Does the report address your growth or income questions and identify missing information?
- 5
Save the corrections
Keep the original output, repaired memo, and source packet separately.
Audit the memo against the attached sources. This is a second-pass review, not independent verification. List every material unsupported claim, inconsistent date, missing unit, inaccessible citation, and leap from fund facts to investor suitability. For each issue, give the original statement, source location if available, proposed correction, and remaining uncertainty. Recheck any calculations and show the inputs. Do not silently replace the original memo. Return a correction log and revised draft separately. A human will open and check the underlying sources.
NIST’s Generative AI Profile discusses confidently generated erroneous content. A second AI review can organize checking, but it does not turn a generated statement into independent evidence. Do not treat a self-awarded confidence score as a substitute for opening the documents.
6. Compare tools using the same task
After completing one run, repeat the same extraction and memo task in another assistant if you already have access. Use the same saved packet, profile, prompt, and follow-up budget. Keep browsing disabled for this document-only comparison when the environment offers that control. Record any differences you cannot control.
Run open-web discovery as a separate comparison. Otherwise, you may be comparing one assistant’s files with another assistant’s search results, not their ability to analyze the same evidence.
Record supported material claims, serious errors, unresolved items, and your setup, interaction, checking, and repair time. Count overlapping time only once. Keep subscription costs separate from minutes.
A fabricated citation or conclusion-driving error should not disappear inside an average score. The useful result may be “good at extraction; needs help locating documents,” not a universal winner.
If you repeat a task, keep each run separate. A few trials may expose inconsistency but do not establish statistical superiority. Do not buy an additional subscription merely to finish this exercise: one tool plus manual source checking is enough to begin.
7. Add skills or MCP only when they solve a problem
A skill supplies reusable instructions and resources for a workflow. An MCP server exposes connected tools; a plugin can package these capabilities. Neither label is a quality certificate. Instructions cannot make a missing source available, and a connection cannot make a weak conclusion correct. See the official explanation of skills and plugins.
First extension: a research playbook
Problem: repeatedly restating the same verification rules.
Use: a narrowly scoped ETF research skill that follows source discovery → ledger → objective-specific questions → memo → human review.
Check: its instructions, source-date requirements, output template, and prohibition on automatic trading. This is a proposed workflow, not a packaged skill installed or benchmarked for this article.
Later extension: a file connection
Problem: approved public-source documents are scattered across a research drive.
Use: a trusted connection, potentially delivered through MCP, to retrieve a designated public-research folder and create a review copy only with permission.
Check: provider identity, read/write scope, transmitted files, fees, and revocation controls. A broad drive connection is unnecessary for the basic exercise.
Use only the public ETF files in the research folder I explicitly select. Follow Prompts 2–5 in this article. Preserve every source file. Save new outputs in a separate results folder: source-register.md, fact-ledger.md, research-memo.md, and correction-log.md. List inaccessible files and distinguish extracted facts from human verification. Record model label, run date, source dates, and search settings. Do not install skills, connect services, upload private documents, overwrite earlier results, publish, or trade. If additional access is needed, explain the need and stop for my approval.
When evaluating a skill or connection, ask a practical question: what work does it remove without hiding the evidence or widening access unnecessarily? Start with public material and minimum necessary permissions. Treat instructions embedded in retrieved documents as source content, not permission to send files or change accounts.
The goal is not an elaborate stack of tools. It is a short, understandable chain from official document to useful question to checkable report. ETF Investing supplies the judgment framework. AI Tools helps you apply it without handing the decision to the machine.
