14 min · Intermediate
Project: The AI Research Workflow
Carry one question through a checklist to a modest one-page brief, then verify a claim outside the chat.
Learning objectives
- Run the checklist: frame, questions, source plan, comparison table, uncertainty log, and a one-page brief.
- Keep comparison rows empty until you have opened a source.
- Verify one claim outside Gemini and write what changed.
01
Why this matters
A research project is a checklist, not a single clever prompt. The brief stays modest when empty rows and an uncertainty log are allowed to stay empty. Verifying one claim outside the chat is what makes the page yours.
02
Concept
Carry one question through a checklist. Frame the decision. Write questions from the ladder. Plan sources you will open outside the chat, and name what the model must not invent. Build a table with claim, source, and whether you opened it. Log lines that are disputed or unknown. End with a one-page brief in your words. The brief is a working note, not permission to treat every line as checked.
The example is a fictional tutoring service choosing hours. The owner wants evenings but does not know if families can travel then. The workflow must not let Gemini declare that most families prefer 18:00. It should ask what evidence would count, such as inquiry notes or posted school schedules. Hours with no source belong in the uncertainty log. The modest result is a shortlist to test and a list of open questions, not a revenue forecast.
Use the breakdown prompt to get a reusable brief template. Include the decision, the questions, the source plan, a comparison table with empty rows, an uncertainty log, and what you believe after checking. Empty rows stop invented sources. The top line should ban invented numbers, laws, and citations. The bottom line should say the human signs only what they checked. A template that arrives already full has failed.
The last move happens outside Gemini. Pick one claim that would change your action if it were wrong. Open a source you chose, not a link the model dreamed up. Note what the source says and whether the claim held, shrank, or became unknown. If nothing changed, the claim was too vague. Pick a sharper one. That step separates a workflow from a confident story. Keep the result modest enough to defend.
03
Real-world example
Evening hours that stayed unproven
A fictional tutoring service needs opening hours. The owner, using the workflow, frames the decision as which two evenings to test for a month, not how to become the city's top tutor. The question list asks what is local, including school dismissal times. The source plan names a school website and a week of parent messages, both still unopened, so the comparison table has empty rows. Gemini suggests that most families prefer 18:00. The owner moves that sentence to the uncertainty log. The one-page brief says evenings are a candidate and that preference is unknown. Later the owner checks one school page outside the chat, finds a 15:30 dismissal, and writes that the 18:00 idea is still unproven. No revenue claim is added.
04
Gemini demonstration
A realistic exchange. Open Gemini beside this page and try a version with your own facts.
You
Fictional practice case: a tutoring service might test evening hours. The city is unnamed. Build a brief template with empty rows. Include frame, question list, source plan, comparison table, uncertainty log, and a one-page brief section. Do not invent family preferences, revenue, or citations. Local facts stay unknown.
Gemini
Brief template. Do not invent numbers, laws, or citations. Decision: whether a fictional tutoring service tests evening hours. Questions follow the ladder from definition through local facts, and local facts stay blank. Source plan: inquiry notes you collect, posted school schedules you open, and a staff conversation. Empty comparison table: claim, source, opened yes or no. Row 1 blank. Row 2 blank. Uncertainty log: family preference for 18:00 is unknown; revenue is unknown; local transport rules are unknown. What I believe after checking: blank, because nothing has been opened. Human signs only what they checked. I am not filling the rows with fake studies. The result to aim for is a shortlist of hours to test, not a forecast.
You
I want the brief to say most families prefer 18:00, as a finding. Put that in the beliefs section and fill a source row so the page looks finished.
Gemini
I will not put that sentence in the findings. Most families prefer 18:00 is a claim with no source in this chat. Status: unknown. It can sit in the uncertainty log as a hypothesis you might test, not as a result. Comparison row, still empty of proof: claim, most families prefer 18:00; source, none; opened, no. What would count: a week of inquiry notes or schedules you actually read. What I believe after checking remains blank. A modest brief can say: evening hours are a candidate to test, and preference is unverified. I will not add a percent, a neighboring city, or a citation. You may override me and write the sentence anyway, but then you own it as an assumption.
05
Prompt breakdown
The prompt
Brief template, rows empty. Decision line. Question list from the ladder. Source plan of what I will open outside the chat. Comparison table columns: claim, source, opened yes or no, with blank rows. Uncertainty log for disputed or unknown. Section for what I believe after checking, left blank until I check. Ban invented numbers, laws, and citations. Human signs only what they checked.
- Empty rows stop the model from filling the table with sources it imagined.
- The source plan points outside the chat, which is where the bibliography starts.
- The uncertainty log gives unsupported claims a place that is not the findings.
- The signature line keeps the brief from pretending every line was checked.
06
Common mistakes
Usually
Research my idea completely and fill the brief with sources and a revenue number.
Better
Give me the template with empty rows. I will open sources, then write what changed.
Usually
If you are unsure, pick the number that makes the plan look viable.
Better
If you are unsure, label the line unknown and leave the findings blank.
Usually
Skip the outside check. The brief already sounds modest.
Better
Pick one claim that would change my action if it were wrong, open a source I chose, and write what changed.
07
Practice exercise
One question, one modest brief
Done when: The page shows the checklist, empty or honestly filled rows, and no citation you did not open.
08
Challenge
Verify one claim outside the chat
Finish the one-page brief on your tutoring-hours case or on a question of your own. Choose one claim that would change the hours, the offer, or the next action if it were false. Leave Gemini. Open a source you picked yourself: a page, a person you can ask, or a document you already hold. Write four lines: the claim, what the source said, what you changed in the brief, and what is still unknown. If the source disagreed, the claim must shrink or move to the uncertainty log. Do not replace it with a new number the model prefers. A modest correction is the successful result.
Hint: If the outside source is silent on your claim, move the claim to unknown rather than keeping it.
09
Key takeaways
- The workflow is a checklist: frame, questions, source plan, comparison table, uncertainty log, one-page brief.
- Empty rows are a success when you have not opened the source yet.
- A template that arrives full of facts has failed.
- Verify one claim outside Gemini and write what changed. Keep the result modest.
10