Case study
recommendation agent.
A program-recommendation agent for a veterans education site, working from a database of 100+ schools and 1,000+ programs. The prototype was built and it worked, but compliance judged that naming specific programs crossed the line, so it never launched.
- Role
- Prompt design in the OpenAI console
- Type
- AI agent (program recommendation)
- Year
- 2025
- Status
- Built and working · not launched (DoD/VA rules)

The spark
I wrote the prompts for this agent at my previous job, a marketing agency for universities. One of its sites serves veterans looking for education programs, and its database holds more than 100 schools and 1,000 programs. That is more than anyone can browse, so the idea was an agent that takes a veteran's situation and narrows the database down to programs that fit.
The problems to solve
- P1
Too many programs to browse
With 100+ schools and 1,000+ programs in the database, browsing doesn't get a veteran very far. They needed the few programs that fit their own situation.
- P2
The rules left a gray area
DoD and VA rules govern how schools and programs are recommended to veterans, but they don't say whether filtering programs by eligibility counts as a "recommendation". Debating that in meetings couldn't settle it.
- P3
The agent's behavior lived in its prompts
The agent ran on the company's shared OpenAI agent, and I shaped it through prompts written in the OpenAI console. Every rule about benefits, eligibility and programs had to hold up in prompt form.
The question the team kept returning to: should the agent name programs at all?
How it works
- 01
School and program database
The agent drew on a database of 100+ schools and 1,000+ programs, so its answers came from the site's own program data.
SolvesP1
- 02
Shared OpenAI agent, set up in the console
The build used the company's shared OpenAI agent. I wrote and revised the prompts in the OpenAI console, which is where the agent's behavior was set.
SolvesP3
- 03
Benefits and eligibility before programs
The team leaned toward a flow that starts with the veteran: which education benefits apply and what they're eligible for. Programs come after that, matched on eligibility.
SolvesP1P2P3
- 04
Two guardrails in the prompts
I wrote two guardrails into the prompts. Programs were matched on eligibility criteria only, with no ranking or endorsement, and every output carried a disclaimer.
SolvesP2P3
Product decisions
A prototype for compliance to judge
Since the rules didn't answer the question, the team built a working prototype so compliance could judge real outputs instead of a description of them.
SolvesP2
Whether to name programs at all
Strict DoD/VA rules made the team question whether the agent should name programs at all. The lean was to circle around it: lead with benefits and eligibility, and keep anything that could read as an endorsement out of the replies. The prototype's review later settled the question.
SolvesP2
Eligibility-only matching, no ranking
Where programs did come up, they were matched on eligibility criteria alone. The list had no order that could read as ranking or endorsing partner schools.
SolvesP1P2
Built with
Where the line sat
The agent never launched and ran on the company's shared OpenAI agent, so there is nothing public to open. This table sums up where the line sat. It is illustrative and is not a transcript.
| Kind of reply | How it reads | Direction |
|---|---|---|
| Explains which education benefits may apply | Information about benefits | Kept: lead with this |
| Covers eligibility for those benefits | Information about eligibility | Kept: lead with this |
| Filters programs by eligibility | Gray area: the rules don't say if this is a recommendation | Tested in the prototype |
| Names specific programs | Over the line, per the compliance review | Dropped |
| Ranks partner schools first | An endorsement | Dropped |
| Calls one program the best choice | An endorsement | Dropped |
Outcome
The prototype was built and it worked, which gave compliance real outputs to review. Their answer was that naming specific programs crossed the line, so the agent didn't launch. It never reached users, so there are no usage numbers.
What I took from it: compliance review belongs before the build, so I now put it in week one.