Your First AI Research Project: Research, Check, and Improve

Complete a Real Research Task You Can Explain and Defend

This project turns the method from the Essential AI Research Tutorial into a finished result.

You will research one practical question, inspect the sources, write a provisional conclusion, evaluate it with the Beforemark Research Check, correct weaknesses, and run the check again.

Estimated effort: 45–75 minutes
What you need: One web-connected AI tool, a browser, and a place for notes
What you will finish with: A short, source-backed research brief you understand

Where you are: Choose → Learn → Complete → Check → Improve → Continue

The Project

Research this beginner-friendly question:

What evidence-based habits most consistently improve sleep quality for generally healthy adults, and which claims remain uncertain or depend on the individual?

Why this topic works:

• It has accessible primary and authoritative sources.
• It includes both well-supported guidance and exaggerated popular claims.
• Sources may emphasize different outcomes or populations.
• It requires a careful conclusion rather than a simple yes-or-no answer.

This is an educational exercise, not personal medical advice. Do not enter private health information or use the project to diagnose or treat a condition. Anyone with persistent sleep problems, possible sleep apnea, medication concerns, or other medical issues should consult a qualified healthcare professional.

Prefer your own topic? You may adapt the project if it is low-stakes and has accessible sources. For your first attempt, the supplied topic is recommended because it tests every part of the method without requiring specialized knowledge.

Step 1: Define the outcome

Copy this into your notes:

I want to identify a small set of sleep habits that are supported by credible evidence, understand where the evidence is limited, and explain the findings accurately to another adult.

Your completion standard is not “AI gave me an answer.” It is “I can explain what the evidence supports, where it is uncertain, and why.”

Step 2: Use these subquestions

Investigate:

1. What does “sleep quality” mean, and how is it commonly measured?
2. What evidence supports keeping a consistent sleep and wake schedule?
3. What evidence supports managing light exposure, especially evening light and morning light?
4. How do caffeine timing, alcohol, and exercise affect sleep?
5. What bedroom or behavioral changes have meaningful evidence behind them?
6. Which popular sleep-improvement claims are overstated, mixed, or highly individual?
7. When should sleep difficulty lead someone to seek professional evaluation?

These subquestions are part of the project design. Do not expand the project into a complete guide to every sleep intervention.

Step 3: Run the first research pass

Paste this prompt into your selected tool:

Help me research this educational question: What evidence-based habits most consistently improve sleep quality for generally healthy adults, and which claims remain uncertain or depend on the individual?

Investigate:

1. how sleep quality is defined or measured;
2. consistent sleep and wake schedules;
3. evening and morning light exposure;
4. caffeine timing, alcohol, and exercise;
5. bedroom and behavioral changes;
6. popular claims with mixed, weak, or highly individual evidence; and
7. signs that warrant professional evaluation.

Prioritize current primary and authoritative sources, including relevant government health agencies, professional medical organizations, systematic reviews, and original research where useful. Use independent sources rather than relying on multiple pages that repeat the same study or guidance. Cite every material factual claim with a direct link.

Separate well-supported findings, reasonable but limited findings, conflicting evidence, and unknowns. Do not diagnose, prescribe treatment, or give a final recommendation yet.

Save the response. It is your first research pass, not your final brief.

Step 4: Select the claims that matter

From the first pass, choose five to eight claims that would materially affect your answer.

Examples of claim types—not preapproved conclusions—include:

• a definition or measurement of sleep quality;
• an effect associated with schedule consistency;
• an effect associated with light timing;
• a claim about caffeine, alcohol, exercise, or the sleep environment;
• a limitation or population difference; and
• a sign that professional evaluation may be appropriate.

Avoid filling the brief with trivia. Every chosen claim should help answer the research question.

Step 5: Inspect and record the evidence

Create this table in your notes:

Key claim	Source link	Source type	What the source actually supports	Limitation or conflict	Confidence
1					High / Medium / Low
2					High / Medium / Low
3					High / Medium / Low
4					High / Medium / Low
5					High / Medium / Low

For each claim:

1. Open the cited source.
2. Confirm the source discusses the same claim, population, and outcome.
3. Find the part that supports or qualifies the claim.
4. Record what it actually shows in your own words.
5. Note important limitations, conflicts, or uncertainty.
6. Assign confidence based on the evidence—not the AI’s tone.

If a link is inaccessible or does not support the claim, replace it. Do not count a citation you could not inspect.

Step 6: Add missing or conflicting evidence

Paste this follow-up prompt:

Review the key claims and sources in the research so far. Identify:

• claims that rely on weak, indirect, commercial, or duplicated evidence;
• credible evidence that conflicts with or limits the current findings;
• missing populations or conditions that affect whether the result applies; and
• one or two primary or authoritative sources that would most improve the research.

Do not simply generate more links. Explain which weakness each additional source addresses.

Open the most useful new sources and update your evidence table. More sources are valuable only when they improve quality, diversity, or coverage.

Step 7: Write a provisional conclusion

Write 200–350 words containing:

1. Best-supported findings: Which habits have the strongest and most relevant evidence?
2. Conditions and limitations: When might the findings not apply?
3. Conflicts or uncertainty: Which claims remain mixed or incomplete?
4. Appropriate boundary: When should someone seek professional guidance?

You may ask AI to organize your verified notes, but use this restriction:

Draft a 200–350 word provisional conclusion using only the verified evidence table below. Do not introduce new claims or sources. Clearly distinguish strong evidence, limited or conflicting evidence, and unanswered questions. Preserve the educational—not medical-advice—boundary.

Read the result and rewrite any sentence you cannot personally explain.

Complete the Beforemark Research Check

Evaluate the work you have actually completed—not what the AI claims it completed.

Score each area:

• Pass: Good enough for this project.
• Improve: A material weakness remains.
• Not checked: You do not yet have enough information to judge.

1. Question clarity

Pass if: The research answers the stated question for generally healthy adults and does not drift into diagnosing or treating individual conditions.

Improve if: The answer becomes a broad collection of sleep tips, changes the population, or never defines the intended outcome.

Your score: Pass / Improve / Not checked
Evidence or note: ______________________________

2. Source quality

Pass if: Important claims are supported mainly by authoritative guidance, systematic reviews, or relevant original research—not marketing pages, unattributed summaries, or AI assertions.

Improve if: A material claim depends on a weak source, inaccessible citation, or source that does not directly support it.

Your score: Pass / Improve / Not checked
Evidence or note: ______________________________

3. Source diversity

Pass if: The evidence comes from more than one independent institution or underlying body of evidence and includes suitable source types for the claims.

Improve if: Several citations repeat the same study, press release, commercial position, or institutional guidance without independent support.

Your score: Pass / Improve / Not checked
Evidence or note: ______________________________

4. Claim support

Pass if: Every claim that materially affects the conclusion has a source you opened, and the source supports the claim with the necessary context.

Improve if: A citation is merely nearby, supports a narrower claim, applies to a different population, or was never inspected.

Your score: Pass / Improve / Not checked
Evidence or note: ______________________________

5. Conflicting evidence

Pass if: The brief identifies meaningful disagreement or explains why apparently different findings are not directly comparable.

Improve if: It presents one smooth answer despite credible conflicting or qualifying evidence.

Your score: Pass / Improve / Not checked
Evidence or note: ______________________________

6. Missing information

Pass if: The brief names important gaps, limitations, or conditions that could change how the findings apply.

Improve if: It implies completeness or certainty while important variables remain unexamined.

Your score: Pass / Improve / Not checked
Evidence or note: ______________________________

7. User understanding

Pass if: You can explain the main findings, strongest evidence, and major limitation without reading the AI response word for word.

Improve if: The conclusion sounds convincing but you cannot explain how the evidence leads to it.

Your score: Pass / Improve / Not checked
Evidence or note: ______________________________

8. Appropriate final conclusions

Pass if: The conclusion matches the strength of the evidence, avoids diagnosis or guarantees, and includes relevant limitations and professional-care boundaries.

Improve if: It overstates certainty, treats association as proof of cause, gives personalized medical advice, or reaches beyond the verified evidence.

Your score: Pass / Improve / Not checked
Evidence or note: ______________________________

Interpret Your Research Check

• Eight Pass scores: Move to the completion standard.
• Any Improve score: Complete the matching correction below, revise the brief, and rerun all eight checks.
• Any Not checked score: Treat it as unfinished. Gather enough evidence to score it.

The goal is not a perfect academic review. The goal is a useful conclusion with no known material weakness hidden behind confident AI writing.

Improvement Workflow: Fix the Weak Area

If question clarity is weak

Rewrite the question and outcome in one sentence. Remove findings outside that scope. Ask:

Which parts of the current response do not directly answer this exact research question, and which required parts remain unanswered?

If source quality is weak

Replace the weakest material source with a primary or authoritative one:

For [claim], find a current primary or authoritative source that directly supports, rejects, or qualifies it. Explain why this source is more suitable than the one currently used.

If source diversity is weak

Trace repeated claims to their original evidence, then add an independent source:

Which of these sources rely on the same underlying study, data, or guidance? Find an independent source or evidence base that tests the claim separately.

If claim support is weak

Narrow, remove, or replace the claim:

Compare this claim with the cited source. State exactly what the source supports, what it does not support, and provide a corrected version of the claim without overstating the evidence.

If conflicting evidence is weak

Search specifically for credible disagreement:

Find the strongest credible evidence that challenges or limits [finding]. Explain whether the difference comes from methods, population, definitions, timing, or a genuine conflict.

If missing information is weak

Identify the gap most likely to change the conclusion:

What important missing variable, population, limitation, or alternative explanation would most affect this conclusion? Find the best available evidence about it.

If your understanding is weak

Close the AI response and explain the answer aloud or in writing. Then compare your explanation with your evidence table. Ask for clarification only on the point you cannot explain:

Explain [specific concept] in plain language, then ask me two questions that test whether I understand it. Do not introduce new claims.

If the conclusion is too strong

Calibrate the language:

Rewrite the conclusion so each statement matches the confidence in the evidence table. Replace guarantees and universal claims with accurate conditions, limitations, or uncertainty. Add no new information.

Revise and Run the Research Check Again

After correcting every Improve or Not checked item:

1. Update the evidence table.
2. Revise the provisional conclusion.
3. Rerun all eight Research Check criteria—not only the weak ones.
4. Repeat the loop if a material weakness remains.

Research is complete when the known weaknesses have been corrected or transparently acknowledged—not when the tool stops generating text.

Completion Standard

Your first AI research project is complete when:

• all eight Research Check areas pass;
• you opened every source supporting a material claim;
• the source table distinguishes evidence from inference;
• the brief acknowledges important conflicts, gaps, and limits;
• the conclusion does not exceed what the evidence supports;
• you can explain the result in your own words; and
• the result stays within the educational and safety boundaries of the project.

Save your final evidence table and 200–350 word brief. That is the finished project.

Final Reflection

Answer these three questions:

1. Which part of the AI’s first response changed most after you checked the sources?
2. Which source or limitation most affected your conclusion?
3. What will you do differently the next time you research with AI?

If your answer to the first question is “nothing,” make sure you actually challenged the initial response. Good research often changes the wording, confidence, scope, or evidence even when the general conclusion remains similar.

Your Next Step: Use the Method on Your Own Topic

You have completed the pilot when you can repeat the method without depending on the sample topic.

Choose one low-stakes question that matters to you. Use the same sequence:

1. Define the outcome.
2. Write a neutral research question.
3. Create subquestions.
4. Set the evidence standard.
5. Run the research pass.
6. Inspect and record the sources.
7. Write a provisional conclusion.
8. Complete the Beforemark Research Check.
9. Improve weak areas and rerun the check.

Need to review the method? Return to the Essential AI Research Tutorial
Want to change tools? Return to the AI Research Journey Hub
Continue learning: Browse Beforemark Guides