POLISCI 290 Capstone Project — Overview

Download this page for your AI (.md)

The project in one paragraph

You write an applied research project: you identify a real-world problem or puzzle, formulate a thesis, argument, or actionable strategy that draws on political science, and support your claims with empirical analysis. You choose the field and the format, from a research paper to a strategy memo or a risk report, the way political science is used in professional settings (see Capstone formats). For graduating seniors, the project is the capstone of the major: it pulls together what you've learned across your coursework.

The capstone is an individual project.

What a good project looks like

• A specific question. Not a topic. "Immigration and elections" is a topic. "Did county-level exposure to ICE raids in 2018 change Latino turnout in the following midterm?" is a question.
• A clear claim. You propose an answer, a strategy, or a solution, and you can say why it holds (the mechanism).
• Grounded in political science. Your argument draws on the scientific literature, not only on news and opinion.
• Real evidence. A concrete measurement strategy, data you can actually get, and an analysis that speaks to your claim.
• Honest about limits. Your claims are only as strong as your evidence.
• Written for its audience. A strategy memo reads like a strategy memo; a research paper reads like a research paper.

Milestones and deadlines

Deadlines are 11:59 p.m. unless stated.

You write the project one piece at a time, with a milestone due every Sunday. The early milestones each draft one section of the final paper. The later ones expand, edit, and polish the full draft.

Only your best 6 milestone grades count. But 3 points are taken off your total milestone points for each milestone you don't submit, so submit all nine.

Office hours. You must attend at least two office hours to discuss your progress: one early in the quarter and one toward the end. Book them through the sign-up link in the syllabus.

How it is graded

Rubrics for each graded component are in Rubrics.

The final paper: format rules

• At least 5000 words, about 20 pages, not counting the bibliography, title page, and figures.
• Times New Roman, 12 pt, double-spaced.
• Any citation style (MLA, Chicago, APA…), as long as it's consistent.
• An AI-use note attached (see AI use and disclosure). The same goes for every milestone.

How the pieces connect

Question ─► Thesis ─► Data Plan ─► Results ─► Rough Draft ─► Edits (theory, empirics, framing, polish) ─► Presentation ─► Final paper
  Sep 27     Oct 4      Oct 11      Oct 18       Oct 25          Nov 1 · Nov 8 · Nov 15 · Nov 22         Nov 30/Dec 2    finals week
  wf 01      wf 02      wf 03       wf 04        wf 05           wf 06 · wf 07 · wf 08 · wf 09            wf 11           wf 12

The first four milestones each become a section of your paper: the question becomes the introduction, the thesis the argument section, the data plan the research design, and the preliminary results the first version of your findings. From week 5 on, you work on one paper that grows from 10 to 20 pages. Nothing you write is thrown away. Workflow 10 covers the two progress reports, which can come at any point in the quarter.

The files in this package

See How to use this guide for how to use them, and how to hand them to your AI tool.

POLISCI 290 Capstone Project Package

Download this page for your AI (.md)

This folder has everything you need to plan and write your capstone project:

• one overview of what the project is and how it is graded;
• twelve short workflows: one for each weekly milestone, one for the progress reports, one for the final presentation, and one for the final paper;
• four supporting docs: capstone formats, rubrics, AI use and disclosure, and working with an AI agent;
• an optional replication starter: one folder for your inputs, outputs, scripts, writing, and visible AI work.

Start with polisci290-00-project-overview.md.

What's inside

How each workflow is built

Every workflow has the same parts:

1. Goal: what this week is for.
2. You produce: the concrete output to submit, or to take to the next week.
3. Steps, each tagged by who does it:
   • You: work you do yourself.
   • With AI (optional): back-and-forth with your AI tool, with a ready-made prompt at that step. Every AI step is optional: you can always do it yourself instead.
   • You decide: the choices that are yours to make, often after an AI step.
4. Check before moving on: a short self-test.
5. Common traps.

The workflows are a suggestion, not a requirement. They show one good way through each week. If a different order, or a different mix of your own work and AI, fits your project better, work out your own. You can even hand this guide to your AI tool and ask it to make a version that follows your own project plan, in the same format. The milestones, deadlines, and lengths stay the same either way.

Using these files with your AI tool

These files are written so you can hand them to an AI model (whichever frontier model you have through Stanford, or the Stanford AI Playground).

1. Open a new chat for each week's milestone.
2. Paste in, or attach, that week's workflow file and your own work so far: your previous milestones and any feedback you got.
3. At each With AI step, paste that step's prompt.
4. Work through it with the AI. Make the decisions yourself.
5. Keep a copy of the chat. You'll need it for your AI-use note.

What the AI does in these prompts. It questions you, suggests options for you to choose from, gives you feedback, and can help with working material such as code and data cleaning. The text you submit is yours. In this course, all assignments must be written in your own words: submitting AI-written text as your own violates the Stanford Honor Code. See docs/polisci290-ai-use-and-disclosure.md for the exact rules.

If your AI tool can work with files

Start with the guide to working with an AI agent. It includes a small folder setup, a copyable first-task prompt, a status page and work log, and a decision record that you own. The replication starter is included in the package; on the website you can also download it separately. No coding or Git knowledge is needed to set up the folder. Using an agent is optional.

Workflow 01 — Revisiting your project question

Download this page for your AI (.md)

When: week 2, early in the week. You already submitted the Project Question (Sun, Sep 27). Feeds: the Working Thesis (Oct 4), and later your introduction.

Goal

You've written a question. Before you build a thesis on it, take one more pass to make sure it's specific, answerable, and motivated. A sharper question now saves you weeks later. This week's work is optional, but most questions get better with a second look.

You produce

• Your question, revised or confirmed: 1–2 sentences, ending in a question mark.
• Its type: positive (what is happening, and why?), normative (what should happen?), or interpretive (what does it mean to the people involved?). See Session 1A.
• Your key variables: the cause and the outcome (or the main concepts), each with a one-line definition.
• One line on what changed, and why. If nothing changed, say why the question holds.

Steps

You · With AI · You decide

1. You Reread your submitted Project Question, and any feedback you got in class or from the instructor. Write down, in one sentence, what you actually want to know.

2. You Check it against Session 1B. Name your key variables. For each, write what it means in your project (the systematized concept) and one way you could observe it.

3. With AI (optional) Stress-test the question. Paste in your Project Question milestone and your notes from steps 1–2:

   Prompt

   You are a research-design mentor for a political science capstone seminar.
   Here is the project question I submitted and my notes on its key variables. Do not rewrite it for me.
   1. Tell me whether my question is a topic in disguise, and why.
   2. Ask me who, where, when, and what outcome, one at a time, until the question is specific.
   3. Tell me whether the question is positive, normative, or interpretive, and whether it assumes its own answer.
   4. Check that my key variables are distinct concepts I could observe, and flag any that are too vague.
   5. Ask me "so what?" twice: who should care about the answer, and what would they do with it?
   6. Suggest 2 sharper versions that go in different directions, and say what evidence each would need.

4. You decide Keep your question, revise it, or combine it with one of the versions. The question has to be one you understand and care about. Check two things: could an answer you dislike come out of it, and is it small enough for ten weeks?

5. You Write down the final question, its type, your key variables, and one line on what changed. Carry them into Workflow 02.

If the AI suggested a framing you adopted, note it in your AI-use note for the Working Thesis.

Check before moving on

• A stranger could read your question and know exactly what you're investigating.
• You can name the question type, and your key variables.
• Each key variable has a definition and at least one way to observe it.
• You can say who would care about the answer, and why.
• The final wording is yours.

Common traps

• Topic, not question. If it has no question mark, it isn't a question.
• Smuggling in the answer. "Why does social media destroy democracy?" assumes that it does.
• Changing the question every week. Revise it now if you need to. After the Data Plan, change it only if the evidence forces you to.
• Taking the AI's version as is. Its suggestions are options, not your question.

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Sessions 1.A and 1.B).

Workflow 02 — The working thesis

Download this page for your AI (.md)

When: week 2, before Sun, Oct 4. Class: Session 2, drawing insights from the scientific literature. Feeds: the Working Thesis milestone, at least 350 words and 5 citations, and later the argument section of your paper.

Goal

Move from a question to an answer: a thesis you can defend, the mechanism that makes it work, and the literature it builds on. It's a working thesis: it can change as your evidence comes in.

The syllabus asks for three things: propose an answer, strategy, or solution; draw insights from political science literature; describe the mechanisms behind the thesis.

You produce

A Working Thesis milestone (at least 350 words, at least 5 citations) with:

1. Thesis: 1–3 sentences. A claim, not a topic: what you argue.
2. Mechanism: the step-by-step process that connects your cause to your outcome, or your premises to your recommendation: how it works. Who are the actors, what do they want, and which institutions constrain them?
3. Literature: at least 5 sources, each doing a job for your argument:
   • allies, which support a step of your mechanism or show how others measured your variables;
   • at least one foil, a competing argument you take on;
   • ideally one template, a study that does something like what you want to do.
4. Scope: where and when your thesis applies.

Steps

You · With AI · You decide

1. You Write a first version of your thesis, in one sentence. State the direction: does X increase or decrease Y? If your project recommends something, state the recommendation and its main reason.

2. You Build the argument yourself first. Pick an approach from Session 2:

   • inductive: start from a concrete case of your topic, and ask what explains it;
   • deductive: start from a general model (actors, incentives, institutions) and ask what it predicts in your case.

   Write the chain: X happens, which leads A to do B, which produces Y. Aim for 3–5 links.

3. With AI (optional) Brainstorm, then challenge, the argument, as in the Session 2 workshop. Paste in your question, your thesis, and your chain:

   Prompt

   You are a political science mentor helping me develop a working thesis for my capstone.
   Here is my question, my one-sentence thesis, and the step-by-step mechanism I have in mind.
   Do not write the thesis for me.
   1. Tell me whether my thesis is a claim or a topic, and whether a reasonable person could disagree with it.
   2. Walk through my mechanism one link at a time. Point out any missing step between cause and effect,
      and tell me which link is weakest.
   3. Brainstorm 3 other mechanisms that could connect the same cause and outcome, with the actors involved.
   4. Suggest the strongest competing argument (a foil), stated the way its supporters would put it.
   5. Ask me where and when my argument should apply, and when it might break down.

4. You decide Settle your thesis and mechanism. Keep what you can defend. Choose the foil you'll take on.

5. You Find your literature. Start with Google Scholar: search by concept (for example, "descriptive representation" AND "turnout"), not by topic. When you find a strong paper, use its bibliography and its "Cited by" list. AI search tools (such as Elicit, or a "deep research" report) can help you find leads.

6. With AI (optional) Map the sources onto your argument. Paste in your thesis, mechanism, and the list of sources with one line on each:

   Prompt

   Here is my thesis, my mechanism, and the sources I found, with one line on what each says.
   1. For each source, tell me what job it could do: support a step of my mechanism, show how to measure a
      variable, serve as a foil, or model the analysis I want to do.
   2. Tell me which steps of my mechanism have no support yet.
   3. Tell me whether I have a real foil, or only sources that agree with me.
   Do not add new sources to my list. If you mention any study, remind me that I have to verify it exists.

7. You Verify every source: it exists, you've read at least the abstract and the relevant section, and it says what you'll use it for. AI tools invent citations.

8. You Write the milestone in your own words. Argue, don't list: "Smith (2020) found X" is a summary; "Because incumbents respond to local media (Smith 2020), a newspaper closure should…" is an argument.

Check before moving on

• Your thesis is a claim a reasonable person could disagree with.
• Your mechanism has no black boxes: every link is a concrete claim about actors or institutions.
• You know which link is weakest.
• At least 5 citations, each doing a job, including at least one foil.
• Every source is verified.
• At least 350 words, in your own words, with an AI-use note attached.

Common traps

• A topic sentence instead of a thesis. "This paper examines X and Y" works just as well for a paper that reaches the opposite conclusion.
• A book-report literature review. Every source should do a job for your argument.
• All allies, no foil. If nobody disagrees with you, either the thesis is too safe or you haven't found the critics.
• Unverified citations. A citation you can't find is a citation you can't use.
• Freezing the thesis. It's a working thesis. If the data plan or the results push back, you'll revise it in week 6.

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Session 2).

Workflow 03 — The data plan

Download this page for your AI (.md)

When: week 3, before Sun, Oct 11. Class: Session 3, developing a data plan. Feeds: the Data Plan milestone, at least 350 words, and later the research design section of your paper.

Goal

Decide what evidence would show whether your thesis holds, where it comes from, and what you'll do with it. The syllabus asks for three things: a concrete measurement strategy; accessible data sources, identified and described; and the mode of analysis.

You produce

A Data Plan milestone (at least 350 words) with:

1. Expectations: 2–4 things you should see in the evidence if your thesis is right. At least one should be something your foil would not predict.
2. Unit and cases: what one observation is (a person, a county-year, a bill, a speech, a country), which ones you'll cover, and why those.
3. Measurement: for each key variable, the indicator you'll use, what it captures, and what it misses.
4. Sources: where each indicator comes from, how you'll get it, and a backup.
5. Mode of analysis: quantitative, qualitative, or mixed, and the comparison that gives you leverage (across units, over time, before and after, within a case).

Steps

You · With AI · You decide

1. You Write your expectations. Start from your mechanism: if each link holds, what would you observe? Write 2–4 expectations, and mark the one that matters most.

2. You Answer: what is one row in my data? (Or one case, one document.) Then: why these cases, and not others? If the honest answer is convenience, say so. Your case-selection readings (Seawright and Gerring; Koivu and Hinze) give you the options.

3. You For each key variable, brainstorm 2–3 ways to measure it, as in the Session 3 "walk and talk". What would you literally see or count if the concept were present?

4. With AI (optional) Find data and check your measures. Paste in your thesis, expectations, unit, and candidate measures:

   Prompt

   You are a research-design mentor for a political science capstone. Here is my thesis, my expectations,
   my unit of analysis, and my candidate measures for each key variable.
   1. Check that my unit of analysis matches the level where my mechanism operates.
   2. For each measure, tell me what it captures and what it misses, and whether the cause and outcome
      are measured independently of each other.
   3. Suggest existing data sources that could contain each measure (public datasets, government records,
      archives, surveys already fielded, media or legislative text). For each, tell me what to check:
      coverage, time period, unit, and how to access it. Tell me that I must verify each one myself.
   4. Ask whether my cases are systematically unusual, and what that would do to my conclusions.
   5. Suggest the comparison that would give me the most leverage over my expectations.

5. You Check your sources yourself. Open each one, find the variables, and confirm the coverage and the unit. Mark each have it, can get it this week, or unsure. Download a sample if you can: ten minutes with the real data will tell you more than an hour of planning.

6. You decide Choose your main source and a backup, your measures, your cases, and your comparison. Prefer existing data. If your project can only work with new data from people (a survey, interviews, an experiment), talk to the instructor before you commit.¹

7. You Write the milestone in your own words.

Check before moving on

• Your expectations follow from your mechanism, and one of them distinguishes you from your foil.
• You can say what one observation is, and why you chose your cases.
• Each key variable has an indicator, and you can say what it misses.
• Every source is have it or can get it this week, and you have a backup.
• You can name your mode of analysis and your comparison.
• At least 350 words, in your own words, with an AI-use note attached.

Common traps

• A unit mismatch. The mechanism is about voters, but the data is at the state level.
• The indicator is the concept. GDP is not "development"; a Polity score is not "democracy". Say what yours misses.
• No comparison. A single case with nothing to compare it to gives you no leverage.
• A dataset you haven't opened. Data access is the most common reason capstones stall.
• Leaping to a survey. Collecting new data takes approvals, sampling, and weeks you don't have.

¹ Research that collects data from people may need review by Stanford's Institutional Review Board (IRB).

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Session 3).

Workflow 04 — Preliminary results

Download this page for your AI (.md)

When: week 4, before Sun, Oct 18. Class: Session 4, implementing empirical analysis. Feeds: the Preliminary Results milestone, at least 350 words, and later the findings section of your paper.

Goal

Run the first round of your analysis, and say honestly what it shows, and what kind of evidence it is. The syllabus asks for two things: implement the proposed data analysis; present and discuss preliminary results.

You produce

A Preliminary Results milestone (at least 350 words) with:

1. What you ran: the analysis from your data plan, in 2–3 sentences, and any change you made to the plan, with the reason.
2. What you found: your first key result, with a table or figure if it helps.
3. What kind of evidence it is: suggestive, descriptive, predictive, or causal (Session 4), and what that lets you claim.
4. What it means for your expectations: supported, contradicted, or unclear.
5. What's next: the analysis or check you'll add in week 7.

Steps

You · With AI · You decide

1. You Look at your data before you test anything. Count your observations, check for missing values, and look at the distribution of your key variables. Write down anything surprising.

2. With AI (optional) Get the smallest thing working first: one table, one chart, one case coded. AI can help you write and debug code, clean data, and set up the analysis you planned. You check what it produces: look at a few cases you know, and make sure the numbers make sense. Save your AI chats: you'll need them for your AI-use note.

3. With AI (optional) Run the main analysis from your data plan. Run what you planned, before you try variations.

4. You Before you look at anything else, write 2–3 sentences on what the result says about your main expectation. Your first reading, before you start building a story, is often the most honest one.

5. With AI (optional) Interpret the result, as in the Session 4 "pass, pair, and critique" workshop. Paste in your expectations, your result, and your first reading:

   Prompt

   You are a skeptical reader of an undergraduate political science capstone. Here are my expectations,
   my first key result (with the table or figure), and my own interpretation.
   1. Tell me what kind of evidence this is: suggestive, descriptive, predictive, or causal,
      and what claim it can support. Point out any causal language my design can't back up.
   2. Tell me whether the result actually speaks to my expectation, or to something else.
   3. Name the 2 most likely alternative explanations for this result.
   4. If the result is null or goes the "wrong" way, help me tell whether it is a real finding
      or a problem with the data or the code. Do not suggest changes designed to make the result significant.
   5. Suggest the single most useful next analysis or check.

6. You decide Settle what you'll claim from this result, and what you'll add next. A result that goes against your thesis is a finding: report it.

7. You Write the milestone in your own words. Report what the evidence shows before you say what it means.

Check before moving on

• The milestone reports an analysis you actually ran, not one you plan to run.
• You've checked the AI-assisted code or output on cases you know.
• You can name the kind of evidence, and your language matches it.
• Your result is connected to a specific expectation.
• Your output (tables, figures, code) is saved and labelled.
• At least 350 words, in your own words, with an AI-use note attached.

Common traps

• Presenting plans as results. The syllabus asks you to implement the analysis.
• Tweaking until it's significant. Adding controls or dropping cases until the result looks good isn't analysis. Report what you planned, and label anything else as exploratory.
• "Causes" from a correlation. Match the verb to the evidence: "is associated with" is often the honest one.
• Trusting code that runs. Code that runs isn't the same as code that's right.
• Losing your output. A result that lives only in a chat window or a console is a result you'll have to redo.

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Session 4).

Workflow 05 — The rough draft

Download this page for your AI (.md)

When: week 5, before Sun, Oct 25. Class: Session 5, framing and selling a project. Feeds: the Rough Draft milestone, at least 10 pages. From here on, you work on one paper.

Goal

Turn four milestones into one paper: pick your format and audience, build the structure, and give every piece of your work a home. The syllabus asks for three things: combine all previous milestones; structure the final project paper; integrate and expand each section.

You produce

A rough draft of at least 10 pages with:

1. A structure that fits your format (see Capstone formats), with a heading for every section.
2. Every earlier milestone in its place, expanded: the question in the introduction, the thesis and mechanism in the argument section, the data plan in the design section, the preliminary results in the findings.
3. A working introduction that states the problem, your answer, and why it matters to your audience.
4. Placeholders for what's still missing, marked clearly (for example, [robustness check, week 7]).

Steps

You · With AI · You decide

1. You Choose your format and your audience, if you haven't yet. Who is the reader: a scholar, a campaign manager, a legislator's staff, a firm's board, an editor? What do they need to decide or understand? Read Capstone formats.

2. You Write your value proposition in two sentences, as in Session 5: what does your project give this audience that they don't already have?

3. With AI (optional) Explore structures. Paste in your four milestones, your format, and your audience:

   Prompt

   You are an editor helping me structure a 20-page applied political science capstone.
   My format is [format] and my audience is [audience]. Here are my four milestones so far.
   1. Suggest 2 different structures for the paper, as section headings with one line each on what goes there.
      One should follow the conventions of my format; the other can depart from them.
   2. For each structure, tell me where each of my milestones goes, and what is still missing.
   3. Suggest 2 ways to frame the project for my audience (what problem it solves for them),
      and say which of my findings each framing puts first.
   Do not write any of the text.

4. You decide Settle your structure and your framing. Then build the annotated structure from the Session 5 workshop: under every heading, write 2–4 bullets on what goes there. Every piece of evidence should have a section. Nothing should be left without a home.

5. You Draft section by section, in your own words. Start with the section you know best. Expand each milestone: the 350-word versions become full sections. Mark what's missing instead of padding.

6. You Write a working introduction last: the problem, your answer in 2–3 sentences, why it matters to your audience, and a road map.

7. With AI (optional) Check your draft before you submit:

   Prompt

   Here is my rough draft and my annotated structure. Do not rewrite it.
   1. Tell me whether each section does the job my structure gives it, and where the draft drifts.
   2. Tell me whether the introduction states my answer, and whether the rest of the draft delivers it.
   3. List what is missing or thin, in order of importance.

Check before moving on

• You can name your format and your audience.
• Every earlier milestone has a place in the draft, expanded.
• The introduction states your answer, not just your topic.
• Missing pieces are marked, with the week you'll fill them.
• At least 10 pages, in your own words, with an AI-use note attached.

Common traps

• Stapling milestones together. Each milestone was written to stand alone; the paper has to read as one argument.
• The format as a costume. A strategy memo isn't a research paper with a different title. Its structure and tone follow its reader.
• Padding to reach 10 pages. Expand what you have; mark what's missing.
• A literature review that runs the paper. Organize the paper around your argument, not around your reading list.

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Capstone Project Guidelines; Session 5).

Workflow 06 — Editing the argument

Download this page for your AI (.md)

When: week 6, before Sun, Nov 1. Class: Session 6, strengthening analytical frameworks. Feeds: the Edit: theory milestone, at least 12 pages.

Goal

Make the theoretical and analytical sections of your paper as strong as they can be: a refined thesis, a complete literature review, and an argument that has faced its strongest counterargument. The syllabus asks for three things: focus on the theoretical and analytical sections; refine the thesis; complete the literature review.

You produce

A revised draft of at least 12 pages in which:

1. The thesis is updated to match what you now know from your preliminary results.
2. The mechanism is stated step by step, with its scope conditions: where and when it applies, and when it breaks down.
3. The literature review positions your project: who you build on, who you argue against, and what you add.
4. The counterargument (your foil) is stated at its strongest, and answered.

Steps

You · With AI · You decide

1. You Reread your thesis next to your preliminary results. Does the thesis still describe what you're finding? If not, revise it now: a thesis that matches the evidence is stronger than one that ignores it.

2. You Find your foil, as in the Session 6 workshop. Write the strongest argument against your thesis in 3–4 sentences, the way its supporters would put it.

3. With AI (optional) Test the argument against the foil. Paste in your thesis, mechanism, and foil:

   Prompt

   You are a demanding political science reviewer. Here is my thesis, my mechanism, and the strongest
   counterargument I found.
   1. Tell me whether I've stated the counterargument at its strongest. If not, state it better.
   2. Tell me what my argument and the counterargument predict differently. Where would the evidence
      tell them apart?
   3. Point out the conditions under which my mechanism would not work (scope conditions),
      and any assumption I'm making without saying so.
   4. Use counterfactual reasoning (Fearon 1991): if my cause had been absent, what does my argument say
      would have happened, and is that plausible?
   5. Tell me the single weakest point in my argument.

4. You decide Decide how you'll answer the foil: with evidence, with scope conditions, or by conceding part of it. Decide the scope of your claim.

5. You Complete the literature review. For each source, ask what it does for your argument. Cut sources that do nothing, and look for what's missing: the strongest critics of your position, and the study your design most resembles. Verify every new source.

6. You Revise the theoretical sections in your own words. Every paragraph should advance your argument.

7. With AI (optional) Review the revised sections:

   Prompt

   Here are the theory and literature sections of my capstone. Do not rewrite them.
   1. Mark every paragraph that summarizes literature without advancing my argument.
   2. Tell me whether each step of my mechanism is supported by a source, a logical argument, or neither.
   3. Tell me whether the counterargument is answered, or just mentioned.
   4. Check that my key terms are used the same way throughout.

Check before moving on

• Your thesis matches what your evidence shows so far.
• Your mechanism has scope conditions and at least one condition under which it fails.
• The strongest counterargument is stated fairly and answered.
• Every source does a job, and every source is verified.
• At least 12 pages, in your own words, with an AI-use note attached.

Common traps

• A straw-man foil. If you can't state the counterargument as its supporters would, you can't answer it.
• Defending the week-2 thesis at all costs. Refining the thesis is this week's job.
• A literature review as a list. "A found X. B found Y." is not a review. Say what each source means for your argument.
• No scope conditions. An argument that works everywhere, always, is usually too vague to be tested.

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Session 6).

Workflow 07 — Editing the evidence

Download this page for your AI (.md)

When: week 7, before Sun, Nov 8. Class: Session 7, empirical validity and evidence robustness. Feeds: the Edit: empirics milestone, at least 15 pages.

Goal

Finish your analysis, find out whether it holds up, and say clearly what it can't show. The syllabus asks for three things: focus on the empirical sections; conclude the data analysis; discuss robustness and limitations.

You produce

A revised draft of at least 15 pages in which:

1. The analysis is complete: every expectation from your data plan has a result, reported as supported, contradicted, or unclear.
2. Two or three robustness checks target the real weak points of your main result, with the standard for passing set in advance.
3. Validity is discussed: the main threats to internal validity (is the relationship real, in your cases?) and external validity (does it travel beyond them?).
4. Limitations are stated, separate from scope: what your design can't show, and why.
5. Tables and figures are clean, labelled, and readable on their own.

Steps

You · With AI · You decide

1. You List every expectation from your data plan and mark each: result in hand, analysis still to run, or dropped (with the reason). Run what's missing first.

2. You decide Choose two or three checks. Pick the things a skeptic would attack first, and decide before each check what would count as passing. For example:

   • Measurement: does the result survive a different indicator for your key variable?
   • Sample: does it survive dropping unusual cases, or a different time period?
   • Rival explanation: does it hold when you account for the foil's explanation?
   • Qualitative evidence: does the account hold if you drop your weakest source, or read the same evidence another way?

3. With AI (optional) Design a secondary empirical test, as in the Session 7 workshop. Paste in your thesis, design, and main result:

   Prompt

   You are a methods reviewer for an undergraduate political science capstone. Here is my thesis,
   my research design, and my main result.
   1. Name the 3 most serious threats to internal validity (confounders, reverse causation, selection,
      measurement), in order of seriousness, for my specific design.
   2. For each, suggest a check I could realistically run this week with my data, and what result would
      count as passing.
   3. Suggest one secondary empirical design that tests the same thesis a different way.
   4. Tell me how far my result can travel: to which other places, periods, or groups, and why.

4. With AI (optional) Run the checks. AI can help you set them up and run them. Whether a check passed is your call, against the standard you set in advance. Save all output.

5. You Build your tables and figures: descriptive labels (not variable names), informative titles, a note on the data source. A reader should understand each one without reading the text. Put secondary results in an appendix.

6. You Revise the empirical sections in your own words. Report what each analysis shows before you say what it means. Then write the limitations: what you can't show, and why it matters.

7. With AI (optional) Review the empirical sections:

   Prompt

   Here are the empirical sections of my capstone, with my tables and figures. Do not rewrite them.
   1. List my 5 most important empirical claims. For each, compare the strength of the claim
      with the strength of the evidence. Flag overclaiming AND underclaiming.
   2. Tell me whether every expectation from my data plan is addressed.
   3. Tell me whether my robustness checks address the real threats, or only easy ones.
   4. Tell me whether any table or figure can't be understood on its own.

Check before moving on

• Every expectation has a result, including nulls and contradictions.
• Two or three targeted checks, each with a pass standard set in advance.
• Internal and external validity are both discussed.
• Limitations are stated plainly.
• Tables and figures stand on their own.
• At least 15 pages, in your own words, with an AI-use note attached.

Common traps

• Checks for show. Three checks aimed at real threats beat ten that aren't.
• Deciding what "passing" means after you see the result.
• Hiding a null. An expectation that isn't supported is a finding. Report it clearly.
• Limitations as an afterthought. A short, honest limitations section makes the rest of your claims more credible.
• Raw output as tables. Console printouts don't belong in the paper.

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Session 7).

Workflow 08 — The full final draft

Download this page for your AI (.md)

When: week 8, before Sun, Nov 15. Class: Session 8, formulating conclusions and recommendations. Feeds: the Full final draft milestone, at least 20 pages.

Goal

Complete the paper, and get its framing right: an introduction that promises exactly what the paper delivers, and a conclusion that says what your findings mean, for whom, and what to do next. The syllabus asks for two things: focus on the introduction and conclusion; finalize the framing of the project.

You produce

A complete draft of at least 20 pages with:

1. An introduction that states the problem, your answer, your evidence in a sentence, and why it matters to your audience, all within the first page.
2. A conclusion that restates what you found (not what you hoped to find), states the implications, and names what should come next.
3. Recommendations, if your format calls for them (strategy memo, green paper, white paper, risk report), each tied to a specific finding.
4. An executive summary or abstract of about 150–250 words (suggested; it helps you, and it becomes the core of your final presentation).

Steps

You · With AI · You decide

1. You Write the executive summary first, as in the Session 8 workshop: problem, answer, evidence, implications, in about 200 words. If you can't write it, the framing isn't settled yet.

2. You Reread your introduction next to your findings. Rewrite the introduction to match what you actually found. The version you drafted in week 5 almost certainly promises something slightly different.

3. With AI (optional) Check the arc of the paper. Paste in the introduction, the conclusion, and your executive summary:

   Prompt

   You are an editor for an applied political science capstone written for [audience].
   Here are my introduction, my conclusion, and my executive summary. Do not rewrite them.
   1. Tell me whether the introduction promises more, less, or something different from what the conclusion delivers.
   2. Tell me whether my answer appears on the first page, and whether a busy reader would find it.
   3. For each recommendation or implication, tell me which finding supports it, and flag any that go beyond my evidence.
   4. Ask me what the next study should be, and whether my conclusion says so.

4. You decide Settle the framing: the one claim the paper stands on, and the audience it's for. Decide which recommendations your evidence can carry.

5. You Write the conclusion in your own words. Turn results into recommendations only as far as the evidence goes: a correlational finding supports "consider" or "pilot", not "adopt everywhere". Think about scale: would what worked in your cases work at scale (the Session 8 podcast)?

6. You Fill the remaining placeholders, so every section is complete.

7. With AI (optional) Review the full draft for integration:

   Prompt

   Here is my full draft. Do not rewrite it.
   1. Tell me where the paper reads like separate milestones stapled together, and where a transition is missing.
   2. Tell me where the same point is made more than twice.
   3. Check that my key terms and variable names are the same throughout.

Check before moving on

• Your answer is on page one.
• The introduction and conclusion promise and deliver the same thing.
• Every recommendation is tied to a finding, and calibrated to it.
• No placeholders are left.
• At least 20 pages, in your own words, with an AI-use note attached.

Common traps

• An introduction written in October. Rewrite it to match what you found.
• A conclusion that only summarizes. Say what it means and what to do next.
• Recommendations the evidence can't carry. Match the strength of the recommendation to the strength of the finding.
• A new argument in the conclusion. If it's important, it belongs in the body.

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Session 8).

Workflow 09 — Polishing the draft

Download this page for your AI (.md)

When: week 9, before Sun, Nov 22. Class: Session 9, presenting, visualizing, and communicating results. Feeds: the Edit: polish milestone, at least 20 pages, and your final presentation.

Goal

Revise the whole paper in focused passes, from the argument down to the commas, so that the version you submit this week is one you'd be happy to hand in. The syllabus asks for one thing: polish the final draft.

You produce

A polished draft of at least 20 pages that:

1. Has been revised in five passes: argument, evidence, structure, clarity, honesty.
2. Has consistent terminology, citations, and formatting: Times New Roman, 12 pt, double-spaced, one citation style.
3. Has figures and tables a reader can understand at a glance.
4. And, for yourself: a short list of what you'll still change for the final paper.

Steps

You · With AI · You decide

1. You Read the whole paper once, start to finish, without editing. Write down the three things that bother you most.

2. With AI (optional) Review the draft in passes. Paste in your full draft:

   Prompt

   You are a reviewer for an applied political science capstone. Here is my full draft.
   Do not rewrite it. Review it in five separate passes and report findings for each:
   1. Argument: is the central claim clear, and does every section serve it?
   2. Evidence: is every claim supported? Is any result misreported or overstated?
      Is causal language used only where the design supports it?
   3. Structure: are the sections in the right order, with transitions between them?
   4. Clarity: which sentences or paragraphs would an educated non-specialist struggle with? Is jargon defined?
   5. Honesty: what limitations are missing? Does the introduction promise more than the paper delivers?
   Finish with a prioritized list of the 5 most important revisions.

3. You decide Choose which revisions to make, starting with the argument and evidence, before the surface. Fixing commas while a section doesn't serve the argument is wasted effort.

4. You Revise, in your own words. Anything the AI rewrites or restyles that you keep must be disclosed; better, rewrite it yourself.

5. You Work on your figures, using Session 9: one message per figure, a title that states the finding, labels a reader can read. The figures you fix now will go into your presentation.

6. You Do the final mechanical pass: word count, formatting, every citation checked against the bibliography, every source verified, figure and table numbers in order.

7. You Write 3–5 lines on what you'll still change for the final paper, including anything you expect to learn at the conference.

Check before moving on

• Five passes done, argument first.
• Terminology, citations, and formatting are consistent.
• Every source is verified; no AI-invented citations.
• Figures state their finding and read on their own.
• At least 20 pages, in your own words, with an AI-use note attached.

Common traps

• Proofreading instead of revising. Revision means rethinking the argument and the evidence, not just the sentences.
• Fixing everything at once. One pass per problem catches more.
• AI-polished prose you didn't write. It breaks the course policy, and it tends to erase your own voice and argument.
• Over-editing into blandness. Clear is not the same as lifeless.

Milestone requirements from the POLISCI 290 syllabus (Capstone Project Milestones; Expectations and Formatting; Session 9).

Workflow 10 — Progress reports

Download this page for your AI (.md)

When: twice during the quarter. One is planned: you pick the date through the signup survey at the start of the quarter. One is random: names are drawn at the beginning of each session, so be ready from week 2 on. Feeds: the progress reports grade (5%), and the milestone you're working on.

Goal

Use five to ten minutes of the class's attention to get feedback on one decision you're facing. A progress report isn't a performance: it's a chance to discuss a challenge or a doubt, and leave with better options than you came with.

You produce

• One choice: the analytical or empirical decision you want feedback on, stated in one sentence. For example: "Should I measure media access by broadband coverage or by local newspaper circulation?"
• The options: two or three ways you could go, and what each would mean for your project.
• One question for the class: what you most want them to tell you.
• For the planned report: 1–2 slides.
• After the report: a note of the feedback you got, and what you'll do with it.

Steps

You · With AI · You decide

1. You Pick the choice. The best ones are real: a decision you haven't made yet, where the options lead to different projects. Keep a running note of open choices each week, so you're ready for the random draw.

2. You In 30 seconds, say what your project is: the question, and the thesis. Then state the choice and your options.

3. With AI (optional) Sharpen the choice before class. Paste in your latest milestone and the choice:

   Prompt

   I'm presenting a 5-10 minute progress report on my political science capstone to my seminar,
   to get feedback on one decision. Here is my latest milestone and the decision I'm facing.
   1. Tell me whether I've stated the decision clearly, and whether my options are really different.
   2. For each option, tell me what it would gain and cost for my project.
   3. Suggest the one question I should ask the class to get the most useful feedback.
   Do not make the decision for me.

4. You For the planned report, make 1–2 slides: the project in one line, the choice and the options. For the random draw, practice saying it without slides.

5. You Present. Then listen: write down what you hear, especially the objections.

6. You decide After class, make the choice. Write a short note of the feedback and what you decided, and use it in your next milestone.

Check before moving on

• You can state the choice in one sentence.
• You have two or three real options, and one question for the class.
• You can describe your project in 30 seconds without slides.
• After the report, you've written down the feedback and your decision.

Common traps

• A status update instead of a choice. "Here's what I did this week" gets you little feedback.
• A choice you've already made. If you're not open to the answer, the class can't help.
• Too much background. Thirty seconds on the project is enough; spend the rest on the choice.
• Not being ready for the draw. Keep one open choice in your pocket every week.

Presentation guidelines from the POLISCI 290 syllabus (Class Presentations; Presentation Guidelines).

Workflow 11 — The final research presentation

Download this page for your AI (.md)

When: the Final Research Conference, in class, Mon, Nov 30 / Wed, Dec 2. Feeds: the final research presentation (10%) and your final paper.

Goal

Pitch your project to the class: one clear idea, the evidence behind it, and why it matters. Then use the audience's questions to improve the final paper.

The syllabus asks for four things: pitch the project (what's the motivating problem, and why is it interesting?); briefly illustrate your argument, thesis, or solution; illustrate your empirical strategy, data sources, and analysis; discuss and visualize your key results.

You produce

• One-sentence takeaway: what the audience should remember. It must be a claim, not a topic.
• A hook: your first 30 seconds. A surprising fact, a concrete case, or a tension your project resolves.
• A slide plan and a timed rehearsal.
• Q&A prep: the 3 most likely questions, one hard question, and your short answers.

A 10-minute structure (suggested)

That's about 8 minutes: plan for about 80% of your time, because everything takes longer live. Q&A comes after the 10 minutes. Results get the most time: they're what the audience came for.

Steps

You · With AI · You decide

1. You Start from your executive summary (Workflow 08). What is done, and what is still in progress? Present what's done, and be clear about the rest.

2. You Write a first version of your takeaway sentence. If you can't say it in one sentence, the talk isn't focused yet.

3. With AI (optional) Test the takeaway and the outline:

   Prompt

   You are a demanding audience member at a student research conference in political science.
   Here is my one-sentence takeaway, my main findings, and my slide outline for a 10-minute talk.
   1. Tell me whether my takeaway is a claim or a topic.
   2. Suggest 2 alternative takeaways my results could support, and say which is strongest.
   3. Tell me where the outline spends too long on setup and too little on results.
   4. Suggest one visual per slide that could replace text (a chart, a map, a diagram, a document).

4. You decide Settle the takeaway and the slide order.

5. You Build the slides, using Session 9: one idea per slide, few words, large type, a real visual where you can. Use the figures you polished in week 9. AI can help you make charts (for example, by writing the plotting code); the talk itself is yours.

6. You Rehearse with a timer, out loud, twice. Don't read your slides.

7. With AI (optional) Practice the Q&A:

   Prompt

   Here is my takeaway and my speaker notes.
   Ask me the three questions a seminar audience is most likely to ask, one at a time, and react to my answers.
   Then ask me the hardest fair question you can think of.

   Most audiences ask some version of: What about [alternative explanation]? How do you know it isn't [reverse causation / selection / a measurement artifact]? Does this generalize beyond your cases?

Check before moving on

• Your takeaway is one sentence, and it's a claim.
• 5–10 slides, 10 minutes (Q&A comes after), timed in rehearsal.
• Results get the most time.
• You have short answers ready for three likely questions and one hard one.
• You don't present plans as results.
• After the conference: you've written down the feedback you received, and what you'll change in the final paper.

Common traps

• Too much setup, not enough results.
• A literature review on a slide. One sentence of context is usually enough.
• Walls of text on slides.
• A hook that doesn't lead anywhere. It has to set up your takeaway.
• Defensive answers. When a question is hard: acknowledge it, answer what you can, say what you can't, and move on.

Format and content from the POLISCI 290 syllabus (Presentation Guidelines: Final Research Presentation; Session 9), with the length set by the instructor on 2026-09-27.

Workflow 12 — The final capstone paper

Download this page for your AI (.md)

When: after the Final Research Conference, through finals week (exact date on Canvas). Feeds: the final capstone paper (40%): at least 5000 words, about 20 pages.

Goal

Hand in the finished project: one integrated argument, with claims that match your evidence, improved by what you learned at the conference. Most of the work is done: this last round is about the feedback, the claims, and the details.

You produce

• The final paper: at least 5000 words, about 20 pages, not counting the bibliography, title page, and figures. Times New Roman, 12 pt, double-spaced, one consistent citation style.
• An AI-use note (AI use and disclosure).

A default structure (adapt it to your format)

Strategy memos, green papers, white papers, and risk reports usually lead with a summary and recommendations, and put the design in a shorter section or an appendix. See Capstone formats.

Steps

You · With AI · You decide

The paper you submit must be in your own words. AI can help you plan and review it, but it can't write it. Anything it restyles or rewrites that you keep must be disclosed.

1. You Go through your notes from the conference and both progress reports. List every question or objection you heard, and mark each: will address, already addressed, or out of scope (with the reason).

2. You decide Pick the revisions that matter most. The conference questions are a preview of how any reader will push back.

3. You Revise, in your own words. Address the points you picked in the text, where a reader would raise them.

4. With AI (optional) Do a final review:

   Prompt

   You are a reviewer for a final applied political science capstone, written as a [format] for [audience].
   Here is my paper. Do not rewrite it.
   1. Tell me whether my answer is clear on the first page, and whether every section serves it.
   2. List my 5 most important claims. For each, compare the strength of the claim with the strength
      of the evidence. Flag overclaiming AND underclaiming.
   3. Tell me whether the strongest counterargument is addressed.
   4. Tell me whether the paper reads like one argument or like milestones stapled together.
   Finish with a prioritized list of the 3 most important last revisions.

5. You decide Choose which revisions to make, and make them. Calibrate your claims: "causes" only if your design supports it, "is associated with" for correlation, description as description. Don't bury null results.

6. You Final checks: word count, formatting, every citation verified and matched to the bibliography, figures numbered, AI-use note attached.

Check before submitting

• Your answer appears on page one.
• Every claim is supported, and causal language only where the design supports it.
• The strongest counterargument is addressed.
• Feedback from the conference and the progress reports is addressed.
• Every citation is verified; no AI-invented sources.
• At least 5000 words; Times New Roman 12 pt, double-spaced; one citation style.
• The AI-use note is attached.

Common traps

• Ignoring the conference. The questions you got are the best free review you'll ever get.
• Last-minute new analyses. Finish and report what you have; new work this late tends to break what already works.
• Hedging everything, or nothing. Make each claim exactly as strong as its evidence.
• AI-polished prose you didn't write. This breaks the course policy.

Requirements from the POLISCI 290 syllabus (Capstone Project Guidelines; Expectations and Formatting; Generative AI).

Capstone formats

Download this page for your AI (.md)

Pick a field and a format

The capstone is meant to replicate how political science is applied in professional settings. You choose a field and a format that fits your interests. The main options:

If you have a field or format in mind that isn't listed, talk to the instructor at the beginning of the quarter.

What every format has in common

Whatever the format, the project must:

• identify a real-world problem or question;
• apply theoretical and analytical insights from political science;
• use an empirical methodology to support its claims with data-driven evidence;
• meet the same milestones, lengths, and formatting rules (see Project overview).

The format changes who you write for, what you lead with, and how the sections are arranged. It doesn't change how strong the argument and the evidence have to be.

How the formats differ (suggested guidance)

The applied formats (memo, green paper, white paper, risk report) usually move the detailed research design to a shorter section or an appendix. It still has to be there: a reader should be able to check how you know what you claim.

Choosing

1. Start from your question, not from the format. Which reader most needs your answer?
2. Check your evidence. A risk report needs indicators you can track; a green paper needs options you can compare; a research paper needs a design that tests a thesis.
3. Read one real example of your format before the rough draft (week 5): a real think-tank report, campaign memo, or long-form article. Notice what it leads with and how it uses evidence.
4. Decide by the rough draft. You can change format before then; after, it gets expensive.

Rubrics

Download this page for your AI (.md)

Each graded part of the project has a rubric below. Each criterion is scored on three levels: Strong (2 points), Developing (1), or Weak (0). All criteria in a component count equally, so your grade for a component is the points you earn divided by the maximum. For example, a Data Plan scored 2, 2, 1 earns 5 of 6 points.

Applies to every component

• Your own words. All assignments must be written in your own words (syllabus AI policy). Submitting AI-written text as your own is an Honor Code matter, not a rubric deduction.
• AI-use note attached. Every milestone and the final paper need an AI-use note (see AI use and disclosure). A missing note costs 5% of that component's grade.
• Minimum length. Each milestone, and the final paper, has a minimum length (see Project overview). Falling short costs 5% of that component's grade.
• Best 6 of 9. Only your best 6 milestone grades count, and each milestone you don't submit costs 3 points from your milestone total.

Milestones (5% each)

Syllabus: each criterion below is one of the milestone's stated expectations.

Project Question (250 words)

Working Thesis (350 words, 5 citations)

Data Plan (350 words)

Preliminary Results (350 words)

Rough Draft (10 pages)

Edit: theory (12 pages)

Edit: empirics (15 pages)

Full final draft (20 pages)

Edit: polish (20 pages)

Progress reports (5%)

Syllabus: focus on the weekly milestone or one aspect of the project; discuss one analytical or empirical choice and elicit feedback. 5–10 minutes with Q&A.

Progress reports are graded on completion: give both of your reports. The table below isn't scored. It describes what a strong report looks like, so use it to prepare.

Final research presentation (10%)

Syllabus: pitch the project; illustrate the argument; illustrate the empirical strategy; discuss and visualize key results. 10 minutes plus Q&A, 5–10 slides.

Final capstone paper (40%)

Syllabus: identify a real-world problem or question; apply theoretical and analytical insights from political science; employ an empirical methodology to support claims with data-driven evidence. At least 5000 words.

AI use and disclosure

Download this page for your AI (.md)

The rule (syllabus, verbatim)

Depending on the context, artificial intelligence (AI) can either enhance and impede learning. There are times when these tools might provide us with new ideas and understandings (e.g., brainstorming or getting feedback). At other points, practicing skills and synthesizing ideas on our own will be crucial for the learning process. Thus, being mindful of when to use these tools will help us navigate the complexity of interacting with AI technologies in the classroom and beyond.

In this class, we will use AI tools, including ChatGPT, that harness large language models as pedagogical opportunities for learning and teaching in the course. Doing so will align with one of our course goals, which pertains to the evaluation of arguments and justification with evidence.

For this class, all assignments should be written in your own words. Just as you cannot take credit for others’ writing in your assignments, you cannot use paraphrasing software (“spinbots”) or AI writing software (like ChatGPT) and submit the output as your own. Doing so in this is a violation of the Stanford Honor Code.

On top of the syllabus rule, every milestone and the final paper need an AI-use note (template below). A missing note costs 5% of that component's grade.

Everything below is practical guidance on following these rules. If anything here seems to conflict with the syllabus, the syllabus wins.

What that means for each part of the project

Suggested practices

We recommend these, but they're suggestions, not requirements. The binding rules are the syllabus policy quoted above and the AI-use note.

Human first, then AI

1. Pass 1, you. Make a real first attempt in your own words: a question, a thesis, a data plan, a draft.
2. Pass 2, AI with control. Then use AI to challenge, clarify, or extend it.
3. Keep the thread. You should always be able to say what was yours before AI, and what changed after.

This makes your AI-use note easy to write, and it protects the learning the capstone is for.

Privacy

Don't paste private, confidential, or identifying material into AI tools. That includes raw interview notes, non-anonymized data, and classmates' unpublished work.

Verification (non-negotiable in practice)

• AI invents citations. Before any source goes into your work, check that it exists, that it's cited accurately, and that it says what you use it for.
• AI isn't evidence. Don't cite an AI system for facts about politics, law, policy, or public opinion.
• You own the errors. Anything under your name is your responsibility, including mistakes you didn't catch.

What to disclose

Disclose when AI meaningfully shaped:

• your question, thesis, argument, or structure;
• your analysis, code, or data;
• your source search;
• your figures or slides.

No disclosure needed for spell-check, grammar fixes, formatting, or a synonym look-up that leaves your phrasing intact.

When in doubt, disclose.

AI-use note template (suggested format)

Attach one to every milestone and to the final paper: a short section at the end of your document works. If you didn't use AI for a submission, say so in one line.

AI-USE NOTE — [milestone name / Final paper]

Tools used: [e.g., model name and version, Stanford AI Playground]

What I used AI for:
- [Task] — [what the AI did] — [what I kept, changed, or rejected]
- [Task] — ...

What is entirely mine: [e.g., the question, the thesis, all the prose]

Verification: [e.g., all 7 sources checked on Google Scholar; code output checked against 20 cases by hand]

Example entry: "Stress-testing my thesis: the AI suggested that local media closures might matter through party recruitment rather than voter information. I kept my mechanism, but added the recruitment pathway as a rival explanation in section 3."

When an AI tool can edit files or run code

Use the optional guide to working with an AI agent and its replication starter. Agree on one small task, keep the work visible, and record your own research decisions. The agent must pause before making a choice that affects your evidence or argument. You should be able to explain the source → steps → result chain and check important results yourself. These practices support the course rules above; they do not add a graded requirement.

Access

You have access to frontier AI models through Stanford, and to the Stanford AI Playground. If cost or access is ever a barrier for a particular task, contact the instructor.

Sources: POLISCI 290 syllabus ("Generative AI", "Academic Integrity"), quoted verbatim where marked.

Working with an AI agent: a small, visible research folder

Download this page for your AI (.md)

An AI agent is an AI tool that can take several steps for you, such as reading files, editing code, and running an analysis. That makes it useful, but also makes it easy to lose track of what happened. The aim here is simple: you direct the research; the agent helps carry out tasks you understand and approve.

This is an optional setup, not an extra assignment or a requirement to use AI. The course AI-use rules still apply: submitted prose must be your own, sources and results need checking, and meaningful AI use needs disclosure. The syllabus and Canvas remain authoritative.

1. Start with one folder

Download the starter folder, unzip it, and rename replication-starter to something like my-capstone. You can also copy the replication-starter/ directory from the repository. Keep the whole project together:

my-capstone/
  README.md                 Your project map and instructions for repeating the work
  AGENTS.md                 Rules you give every AI agent
  input/                    Original data, documents, and source notes
  output/                   Generated tables, figures, and prepared data
  scripts/                  Code that turns inputs into outputs
  docs/                     Your writing and your research decisions
  agent-workspace/           Agent status, work log, and temporary work

The .md files are plain text with headings. You can read and edit them in a text editor; you do not need a special app, terminal commands, or Git to use this folder. GitHub is just where you get the starter. You do not need to upload your project to GitHub.

Replication means someone can follow your instructions and repeat the steps that produced a result. A folder helps only if you also record where the evidence came from, which steps used it, and how you checked the result. For a qualitative project, inputs can be public documents, scripts can be documented search or coding procedures, and outputs can be evidence tables. Do not add code just to fill a folder.

Set it up yourself first (about ten minutes)

1. You: In README.md, write your question, your current expectation, and the next small task in your own words. It is fine to mark things undecided.
2. You: Put permitted source files in input/. In input/README.md, record where each came from, when you got it, and what it covers. Keep originals unchanged. Do not put confidential, private, or identifying material into an AI tool or a public repository.
3. You: Read AGENTS.md. These are the agent's ground rules. Do not ask the agent to weaken them so it can finish faster.
4. You: Start docs/decisions.md with the research choices you have already made and why. Leave open choices pending.
5. With AI (optional): Open the project folder in your tool. Explicitly ask it to read AGENTS.md and the project map. If it cannot access files, attach or paste those files and maintain the tracking pages yourself. Do not assume a particular tool automatically reads an instruction file.

Instructions are not a technical lock. Limit the tool's file access and permissions where available, read its proposed actions, and check the files it actually changed. Start with one agent and one small task at a time.

2. Keep research choices with you

The agent may explain alternatives and their consequences. It must not select an answer on your behalf, treat silence as approval, or turn a suggestion into an approved research choice.

“Clean the data” is too broad. A choice can look technical and still change the research: treating a blank as zero, deleting an outlier, changing a date range, or selecting a merge rule can change your findings. If the task is unclear, the agent should stop and explain the choice in everyday language.

For a pending choice, ask for: the question, the available options, what each changes, and what evidence would help you choose. Discuss it with the instructor if needed. Then write your choice and reason in docs/decisions.md, in your own words, and tell the agent which recorded decision to implement. A general “go ahead” does not authorize new research choices discovered later.

The agent may draft code, technical documentation, and tracking notes. It may give feedback on your writing. It must not write, paraphrase, or restyle the assignment prose you submit as your own.

3. Give one bounded task

Use this prompt after you have made your own first attempt:

Read AGENTS.md, README.md, docs/decisions.md, and agent-workspace/status.md.
My goal for this session is: [one small task].
The research decisions I have already made are: [decision IDs, or none yet].
You may work with these files: [names].
The expected result is: [one checkable output].

Before editing or running anything, explain your plan in plain language:
what you will read, what you will change, what the result will tell me,
and how we will check it. Put that plan in the status page and wait for
my approval. Flag any research choice I still need to make.
Do not write my assignment prose. Do not make research choices for me.

A good first task: “Using input/county-data.csv, list the columns, count rows and missing values, and save a report to agent-workspace/data-check.md. Do not drop, fill in, or recode anything.”

After reading the plan, approve its exact scope or revise it. The agent can then complete the agreed mechanical steps without asking about every line of code. It must pause again for a new research choice, an unexpected problem, or work beyond that scope.

4. Monitor three small records

The status page should say planned, working, waiting for you, or ready for your check, with the last update time. It should name the actual task and files. “Done” without a check is not enough. The agent updates it before starting, when it pauses or changes course, and when it finishes an agreed task. These are written records, not a live activity monitor: an old status does not prove an agent is still working.

Keep the log short: one entry per bounded task, plus an entry when the plan changes or an attempted run fails. Save useful prompts or chat exports in agent-workspace/ when permitted; record the tool/model and enough detail to explain its contribution. You do not need a transcript of every keystroke.

Example: The agent finds 12 counties with missing outcomes. It records the count and pauses, leaving the data unchanged. You decide to exclude those counties for a particular comparison, explain the limitation, and record D-002. The agent implements D-002 in a script, reports the row counts before and after, and points you to the affected table. You check several counties yourself before using the table. It does not try other exclusions to obtain a preferred result.

5. Be able to explain the whole chain

Before accepting an output, close the chat briefly and answer these in your own words:

1. What question is this step helping me answer?
2. Where did the evidence come from, and what does one row, case, or document represent?
3. What did we do to it, which observations changed, and which choices were mine?
4. Which script or written procedure produced this table or figure, and what checks did we run?
5. What does the result support, what does it leave uncertain, and what can it not establish?

You do not need to explain every line of code. You do need to understand the main steps, the assumptions, and why a check would catch a plausible mistake. Ask for a small worked example if you cannot explain a transformation. If you cannot explain an output or its limits, keep it marked not yet checked and do not use it as evidence in your submission yet.

Have the agent end each task with: what changed; the input → steps → output path; checks actually run and their results; uncertainties; decisions needed; the next small step. “The code ran” does not establish that the analysis is right. Spot-check source records, totals, joins, coding, and figure labels yourself, as appropriate. Verify sources by opening and reading them.

6. Make the result repeatable

Before a milestone, update the project README.md so another person can:

• Find the inputs, their sources, and any access restrictions.
• See the script or procedure order and the exact commands or manual steps, starting from the project folder.
• Identify the software, versions, and extra packages needed. Record a fixed random seed if the procedure uses random draws.
• Connect each important result to its input files, script/procedure, and approved decisions.
• Repeat the work into a separate, empty output folder and compare the expected results, without changing the originals. Record what was actually rerun and any differences.

For a manual evidence table, document the document list, search terms, inclusion and coding rules, and source page references instead. A restricted source may require access instructions rather than a shared copy; do not claim that someone without that access can fully reproduce it.

Write your milestone in your own words and use the work log to prepare your AI-use note. The folder and logs support your explanation; they do not replace your judgment or your responsibility for the work.

Going further: plan your project with an agent

Once your folder is set up, you can also use an agent to plan the whole project before you start the analysis.

• Read the approach. The Research Guide (PDF) explains it in three moves: decide what the finished project must contain, list what you already have, then plan small tasks that each come with a check.
• Give your agent the steps. The research-planning skill turns that approach into steps your agent follows. Download research-planning.zip and unzip it into your capstone folder, next to AGENTS.md. It comes from a larger public system, ai-os-public, if you are curious.
• Start with one prompt. Paste this into your tool: Read AGENTS.md, then read research-planning/SKILL.md and follow it to help me plan my capstone. Start with stage 1.
• Keep the choices yours. You make every research choice and write every word you submit. The course AI-use rules still apply. This step is optional.