Use case I: Event recap generator
The same files and the same basic task can produce very different results depending on how you work with the agent. In this case, you will try the task yourself, verify the output, and compare three different build processes.
1. The business task
After every case competition, the club president sends a standard debrief — same sections, same tone, same length, every time. The problem: the person who takes notes is never the person who writes the recap, and turning messy bullet points into the finished document takes two to three hours.
Your task is to use the available files to generate the Fall 2025 recap in the same style and format as the previous approved recap.
2. Data available
- inputs/competition-invite-fall-2025.docx — official invitation from the organizers. Download
- inputs/notes.txt — raw notes from the event. Download
- references/competition-recap-fall-2024.docx — last semester's approved recap. This is the reference for format and tone. Download
3. Try it yourself
Start by exploring the three source files so you understand what information is available and have a rough sense of what the finished recap should look like.
Then generate the Fall 2025 recap using whatever workflow you are comfortable with. Feel free to iterate or try different approaches. Do the best job you can.
When you're finished, keep your output. In the next step, you'll verify it against a checklist.
4. Verify the output
Now compare your output against the reference and source files. Use the checklist below.
| Check | Questions to ask |
|---|---|
| Facts | Are the date, location, team, results, and other key details correct? |
| Completeness | Are all expected sections present? Are there extra sections that do not belong? |
| Format | Do the headings, tables, spacing, and bullets match the reference? |
| Tone | Is the writing professional, positive, and consistent with the previous recap? |
| Content judgment | Did the AI include details that are true but unnecessary or inappropriate for the final recap? |
| Unsupported information | Did it guess, embellish, or add anything that is not supported by the files? |
| Missing information | When information is unavailable, does the document flag the gap or quietly work around it? |
Does your output pass all of the verification checks? Are you confident about the document (e.g., no false or missing information)? What do you think could be improved?
5. Three ways to build the same recap
Now let us take a look how different workflows could generate different outputs. Three attempts:Attempt 1 · Browser chat + one prompt
In the first attempt, I uploaded the three files to Claude in the browser and used a single prompt:
I am generating a recap for the fall 2025 case competition.
Here are the relevant notes, the invitation to the case competition
(with details about date, location, etc.), and a reference recap
from fall 2024.
Generate the 2025 recap for me.Here is the output. It looked plausible at first glance, but the verification checklist exposed several problems.
What went wrong
- It included unnecessary details from the notes. For example: “Phones were collected during the prep period.” The fact is supported by the notes, but it does not belong in a polished club recap.
- The tone could sound defensive. The recap says the team “felt rushed at the end and didn't get a full run-through before presenting.”
- Some language was more informal and evaluative than the reference. For example: “we didn't, and it showed in how hard it was to read the judges” and “Michigan's opening was noticeably sharper than ours.”
- The structure and formatting did not fully match the reference. It added a separate “The case” section and used different formatting choices.
Claude had the source files, but many of the rules for a good recap were implicit. It had to decide for itself what to include, how to frame weaker performance, whether to name people, and how closely to imitate the reference.
Attempt 2 · Project folder + Plan mode + iteration
In the second attempt, I put the same files in a project folder and opened it in Claude Code. I switched to Plan mode and gave Claude the same task:
I am generating a recap for the fall 2025 case competition.
Here are the relevant notes, the invitation to the case competition
(with details about date, location, etc.), and a reference recap
from fall 2024.
Generate the 2025 recap for me.Because Claude could inspect the project before building, it asked clarifying questions during planning — for example, what the consulting club should be called. I answered those questions, let it generate a first draft, and then iterated on the result.
For example, when the styling did not match the reference, I asked:
Compare the new recap with the reference document side by side.
List every inconsistency in headings, tables, spacing, bullets,
and overall formatting. Then fix those inconsistencies.Here is the output. It was much closer to the reference visually. But the deeper content and tone problems were still there because I had never explicitly defined those rules.
What still went wrong
- It added unsupported or overly specific judgments. The results table says: “Jordan opened strong; delivery was our clearest differentiator.”
- It emphasized individual team members in ways that did not fit the desired institutional tone. For example: “Marcus handled the financials calmly”.
- The tone remained conversational and somewhat self-critical. For example: “we left points on the table” and “the question exposed that our number was not as well supported as the rest of the recommendation.”
This output is not necessarily wrong. But a closer reading of the 2024 recap suggests that it is not quite ideal: the reference never names judges in its critique and is written more for an outward-facing audience, emphasizing broader learnings and what the team should do next rather than detailed self-reflection.
Attempt 3 · Project folder + Project rules + iteration
The third attempt used the same project-folder setup, but before generating the recap, Claude was asked to interview the user and create the project's rules.
I am generating a recap for the fall 2025 case competition.
The folder contains the relevant notes, the invitation to the case
competition (with details about date, location, etc.), and a reference
recap from fall 2024.
I would like to generate a recap in the exact same style/format as
the reference recap.
Before we write anything, ask me the questions you'd need answered
to write a good CLAUDE.md — the file that tells you the rules for
this project every time you open it.Claude's questions surfaced requirements that had never been made explicit. For example:
- What should happen when important information is missing?
- Should judges or competing teams be identified by name?
- How closely should the output match the reference?
- What tone should the recap use?
- What output format is required?
- Which details from the raw notes should be omitted?
After answering the questions, I reviewed and approved the generated CLAUDE.md
(download it).
Claude then planned and generated the recap using both the source files and the project rules. Here is the output. It was substantially better. It maintained a more professional tone, matched the reference more closely, and explicitly highlighted missing information instead of quietly guessing or filling the gap.
The workflow and prompt demonstrated here can be improved in many ways, but the final takeaway is simple: good agentic work is rarely about finding one perfect prompt. It is about creating the context, checkpoints, rules, and verification process that help the agent do the right work consistently.