Build an AI-ready portfolio project in four weeks
By First Circuit Editorial · · 3 min read
Saying you're "good with AI" is easy. Showing it takes one small, honest project: a real task, a measured before and after, and a clear account of where the tool helped, where it failed and what you checked. Four weeks of a few hours each is enough.
What makes a project verifiable
- A baseline: how the task is done, and how well, without the tool.
- An evaluation: a small, fixed set of test cases and a way to judge results that someone else could repeat.
- Disclosure: which tools you used, for what, and what you wrote or decided yourself.
- Human review: evidence that you checked outputs and corrected them.
- No private data: use public, synthetic or your own data. Never use client or customer personal information.
Week 1: choose a task and measure the baseline
Pick a task that's common in the work you want and small enough to repeat twenty times. Do it by hand, or in the usual way, for a handful of cases and record how long it takes and how good the results are by a simple rubric you write in advance, such as "accurate, on-brief, needs no edits." Write the rubric down before you see any AI output so you can't move the goalposts.
Week 2: build the AI-assisted version
Design a workflow, not just a prompt: what goes in, what the tool does, and what you check at the end. Save your prompts and settings so the process is repeatable. Keep it simple; a careful workflow with a general-purpose assistant is more impressive than a fragile chain of tools.
Week 3: evaluate honestly
Run your fixed test set through both versions and score them with your rubric. Note time saved, errors introduced and errors caught. Report every failure you observed: cases where the tool was wrong and how you noticed. If you saw none, say what your test set didn't cover, such as unusual inputs or larger volumes, rather than implying the tool is flawless. If you find a test case is ambiguous, don't drop it quietly: record the rubric change, then rerun both versions on the revised set so the comparison stays fair.
Week 4: write it up and publish
Write a one-page case study: the task, the baseline, your workflow, the results table, the failures or test limits, and what you'd change. Add a short disclosure, for example: "AI tool used to draft first versions; rubric, evaluation and final edits by me." Publish it on your portfolio and link it from your resume's projects section.
Make it easy to check. Include your rubric, a sample of test cases and outputs, and the prompts or settings you used, with anything sensitive removed. A reviewer should be able to rerun a few cases and see roughly what you saw. If you can't share something, say why. Finally, ask one person in the field to read the write-up and tell you what's unclear, then fix it before you add it to applications.
Example projects by field
These are project ideas, not completed work. Your results will be your own.
| Field | Task | Baseline and evaluation |
|---|---|---|
| Marketing | Draft product descriptions for a small public catalog | Time per description by hand; rubric for accuracy to specs, brand voice and length |
| Design | Generate alt text for 30 public-domain artworks | Your own alt text as baseline; check accuracy, length and usefulness against WCAG guidance |
| Engineering | Write unit tests for an open-source function | Bugs caught by your tests versus AI-drafted tests; review every test for correctness |
| Operations | Summarize public meeting minutes into action items | Your summaries as baseline; count missed or invented action items |
Common mistakes to avoid
- Choosing a task that's too big. A narrow, finished project beats an ambitious, half-done one.
- Changing the rubric after seeing results. Decide what "good" means first.
- Reporting only successes. Include the failures you saw, or say clearly what you didn't test.
- Using real customer, client or classmate data. Use public or synthetic data instead.
- Overstating results. Report only what you measured. A hypothetical example of the right style: "on my 20 test cases, drafting took about 40% less time," not broad marketing claims.
Also check the terms of any tool you use and of any data you work with, and give credit for open datasets or code you build on.
Talking about it
In an interview, walk through the rubric and what went wrong or wasn't tested first. Explain one decision you made that the tool couldn't, such as revising an ambiguous rubric item and rerunning both versions, or rewriting a section because the facts were wrong. That shows the judgment employers are looking for, whatever tools they use.