Final Project Workbook

The capstone is a real workflow for work you actually have, not a fictional company and not a tour of features. Milestone 1 is the brief: the person you are helping, the decision or artifact, the inputs you may use, and the harm if a draft is wrong. Write that on one page before you prompt. If you cannot name the harm, pick a safer project.

Milestone 2 is the source pack. Gather the notes, specs, or public pages you will allow the model to see. Redact secrets. Label each source. Milestone 3 is the first prompt set: a draft prompt, a critic prompt, and a verification prompt. Run them once on a small slice, not on the whole project.

Milestone 4 is the full draft of the artifact: a research memo, a content piece, a business one-pager, a lesson, or a coded explanation with a check. The artifact must show your edits, not only a paste. Milestone 5 is the evidence pack: the prompts, the claim list, what you changed, and what you refused to ship. A project without that pack is a result you cannot explain.

Leave time between milestones. A single evening cannot hold a source pack and an honest check. If a milestone slips, shrink the artifact rather than skipping the check. Tell a peer the date you will show milestone 4, even if the peer is only you plus a calendar alarm. Deadlines make the capstone different from another practice chat.

At milestone 1, show the brief to someone who does the work, or reread it the next morning. Ask whether the task is real and whether the inputs are allowed. Cut scope until the artifact can be checked. A vague capstone becomes a vague prompt. Fix the brief before you blame the model.

At milestone 2, paste nothing until the labels exist. For each source, write what it is allowed to support. If you are tempted to ask Gemini for sources you do not have, that is a detour, not a milestone. Go find a real source or narrow the claim. The pack should fit the decision, not impress by size.

At milestone 3, save the three prompts in your workbook with a note on the check each one requires. Run them on a slice. Record one failure, such as an invented detail or a tone you would not use. Revise the prompt before milestone 4. Do not discover the failure after you have generated twenty pages.

At milestones 4 and 5, keep a change log. Each entry can be one line: what the model said, what you kept, what you cut, and why. That log is the learning, and it is also your defense if someone asks what you automated. Check the current screen only if a step needs a product feature. Have a paste-text fallback so a missing button does not stop the project.

The reader, user, or decision-maker is named, and the artifact would make sense to them without the chat beside it. The promise of the piece matches the brief. If you drifted into a more exciting project halfway, either update the brief honestly or return to the original. Drift hidden inside a polished file is not completion. The checklist starts with fit, not with flair.

Every factual claim that matters is tied to a source you can open, or it is clearly labeled as your assumption. There are no citations you did not verify, no prices you did not check, and no quotes you cannot find. Gaps are visible. A quality checklist item is failed if the only evidence is that the paragraph sounds finished. Sound is cheap.

The prompts are specific: role, context, task, constraint, and output. They forbid the failure you most expect, such as invented features or fake consensus. A second prompt critiques or extracts claims. You can show the prompts to a classmate without exposing private data. If you cannot, redact a copy for the evidence pack.

Your voice or your professional judgment is visible in the edits. The change log shows at least three deliberate cuts. Sensitive data stayed out. A human other than the model could follow the verification step. If any item fails, the project is not ready for the improvement stage to be skipped. Failing an item is useful information.

The improvement stage is a scheduled revision, not a vibe that you might tidy later. After the quality checklist, pick the two failed or weakest items. Do not start a new project. Rewrite the prompt or the artifact to address those two. Run the check again and record what changed.

Ask Gemini for a critic pass on the artifact using your checklist as the criteria. Tell it not to add new facts. Treat its comments as hypotheses. Accept a comment only if you can see the problem in the file. Reject compliments. The stage is for repair.

Improve the workflow, not only the prose. If verification was rushed, add a claim table and fill it. If the brief was wide, cut a section and say why in the log. If the voice is generic, restore one concrete detail from your source pack. Small structural fixes beat another adjective pass.

Close the workbook with a handover note: what the artifact is, what it must not be used for, which sources it depends on, and what you would do differently next time. That note is the capstone's adult ending. You should be able to reuse the prompts next month without rereading the whole course. If you cannot, the improvement stage is not finished. Done means checked, revised, and explainable.