# Workflow 05 — Doing the work, and documenting it (implementation milestone)

**When:** weeks 5–8, before Sun, Nov 15. **Feeds:** the **implementation milestone**, 800–1000 words (shorter for 4 or 3 units): *design and evidence progress — what has actually been done.*

## Goal

Make real progress, check that it holds up, and **report honestly** what is done, what isn't, and what you learned.

## You produce

An 800–1000-word implementation milestone with:

1. **What you set out to do**: your question and design, in 2–3 sentences.
2. **What you've actually done**: the analysis run, cases coded, or argument drafted. Be concrete.
3. **What you found so far**: preliminary results or a draft argument.
4. **Checks**: how you tested whether it holds up (see step 3 below).
5. **What changed and why**: the design choices you revised since the proposal.
6. **What's left**: the plan to the final submission, including your minimum viable version.

Attachments as relevant: code, a sample of your data, a link to a notebook. Any reasonable format works (a PDF appendix, a link to a shared folder or code repository, a notebook). Make sure the teaching team can open it without having to request access.

## Steps

1. **With AI.** Do the work. Get the smallest thing working first (one case coded, one chart, one working prompt, one section argued), then expand. AI can write code, clean data, run analyses, and draft short results memos for you to read. You check what it produces, and you make the decisions. **Save your AI chats**: you'll need them for your AI-use note.
2. **With AI (optional).** At the end of a work session, ask for a summary you can keep:

   ```
   Summarize this work session in 5 bullet points: what we did, what broke, what I decided and why,
   and what's next. Keep my decisions separate from your suggestions.
   ```

   Read it, correct anything that's wrong, and save it with your project files. It's much easier than keeping a log by hand.

3. **You decide.** Choose **two checks**. Pick the two things a skeptic would attack first, and decide *before* each check what would count as passing. For example:
   - **Empirical:** does the result survive a different measure, dropping unusual cases, or a second coder on a sample?
   - **Theoretical / argumentative:** does the argument survive the strongest objection? Does it hold if you remove your weakest piece of evidence?
4. **With AI.** 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.
5. **You.** Take stock. Go stage by stage (question, design, evidence, write-up) and mark each *done*, *in progress*, or *not started*. Then write 3–5 lines on your key decisions and on what changed since the proposal. Your saved session summaries make this quick.
6. **You.** Write the milestone in your own words.
7. **With AI.** Review your draft:

   ```
   You are a skeptical reviewer for an undergraduate project in a course on AI and politics.
   Here is my implementation milestone draft and my session summaries. Do not rewrite the draft.
   1. Compare what I say I have done with what the summaries show. Flag anything I present as done that is not.
   2. Name the 2 most serious threats to my preliminary results,
      and ask whether my checks actually address them.
   3. Point out where my claims are stronger than my evidence, and where I undersell a real result.
   4. Ask whether my plan to the final submission is realistic, and what my minimum viable version is.
   ```

## Check before moving on

- [ ] The milestone reports what you've *done*, not what you plan to do.
- [ ] Two targeted checks, each with a pass standard you set in advance.
- [ ] Design changes since the proposal are explained.
- [ ] 800–1000 words (or the shorter length for your units), with an AI-use note attached.

## Common traps

- **Presenting plans as progress.** The syllabus asks what has *actually* been done.
- **Checks for show.** Two checks aimed at real threats are worth more than many that aren't.
- **Deciding what "passing" means after you see the result.**
- **Trusting AI output you didn't check.** Code that runs isn't the same as code that's right. Look at the output on a few cases you know.
- **Losing your chats.** Without them, your AI-use note will be vague, and rebuilding what you did is painful.
- **Hiding a null or a failure.** A hypothesis that isn't supported is still a finding. Report it clearly.

---
*Milestone requirements from the POLISCI 158 syllabus, Project Components and Deadlines.*