The challenge
A parent the night before the first day of camp, working out what to pack and what time drop-off actually starts. A camp director with next season’s program to write and a stack of last season’s material to write it from. A development officer with a grant application due and the same organizational background to restate for the fourth time this year.
Two of those people need an answer that is already true. The third needs something that does not exist yet. A nonprofit this size has one team covering all three, and one thing that cannot be allowed to happen: an assistant telling a parent something about a children’s program that the center never said. A community center holds a particular kind of trust, because the people asking are the people whose children are in the building, and an answer that sounds official is treated as official.
The question was never whether to ground the assistants. It was whether every assistant had to be grounded the same way.
The solution
A camp assistant for families. Bounded to the center’s own material, with citations shown on the answer and a persona the center wrote rather than a default one.
A successor camp assistant for the following season. Built fresh rather than by editing the one that had already done its year, bounded and cited in the same way.
A grant drafting assistant and a program knowledge assistant. Both bounded to the center’s own material, so the organizational background the grant tool reuses is the center’s own words rather than a plausible approximation of them.
A curriculum assistant for program staff. The one assistant deliberately allowed to draw on general model knowledge, because drafting is the job and a tool that can only repeat last season cannot help write the next one.
Program documents were uploaded as they stood, and two assistants were pointed at the center’s own website instead. Every assistant that carried a question has a persona the center wrote, and every one of those was revised after it went live. No agentic actions are enabled anywhere on the account, and the assistants reach no registration or membership system.
The capability that decided it
Each assistant is set either to answer only from what the center has uploaded and published, or to draw on the model’s general knowledge. That one setting, decided separately for every assistant on the account, is what this study is about.
Five of the six live assistants are held to the center’s own material. The sixth, which drafts program curriculum for staff, is the exception. Where that choice is made once for a whole account, a nonprofit has to pick between a safe assistant and a useful one, and it will be wrong about one of its three jobs whichever way it picks. Here a parent and a program director can be served by the same platform on the same afternoon under opposite rules.
Anti-hallucination sits alongside it as a setting rather than as a prompt to be re-engineered on every assistant. It is enabled on every live assistant on this account, without exception.
The results
The curriculum assistant for staff has carried 398 questions on 57 days. The camp assistant for families, 304 on 54 days. The successor camp assistant, 56 on 10. The program knowledge assistant, 24 on 3. The grant drafting assistant, 17 on 4. Together, 799 questions on 102 of the 436 days on record.
This is a seasonal organization and the usage admits it. Questions crowd into the weeks around a camp season and thin out to nothing in between. The figure to read is the curriculum assistant’s: 316 questions across its first season, then quiet, then 82 more when the next season came around. Nobody rebuilt it, retrained it or migrated it. It sat there for the better part of a year and was picked up again by the people writing the following summer.
Said plainly, because it matters: these are volume and configuration figures. This account holds no answer-quality measurement of any kind, so nothing here is a claim about whether the answers were right. Every claim on this page is a claim about what the center configured and how much the assistants were used.
Why it worked
The grounding decision is per assistant. One account can hold a strictly bounded assistant and an open one, without a second account and without a second tool.
Refusing came as a setting. Anti-hallucination is on for every live assistant here, which is a choice made once rather than prompt-engineered each time.
Citations went where they belong. On both assistants built around the camp season, where a family is asking, and on neither drafting tool, because a draft is not an answer.
Building one is not a project. Four assistants were created and never carried a question, and that is a sign of a low bar rather than a failure. Nothing is retired and nothing is rebuilt. An assistant goes quiet between seasons and is there when the season comes back.
| Measure | To date |
| Assistants created | 9 |
| Assistants live | 6 |
| Assistants that have carried a question | 5 |
| Live assistants answering only from the center’s own material | 5 of 6 |
| Live assistants with anti-hallucination enabled | 6 of 6 |
| Questions asked, raw count | 799 |
| Days carrying a question | 102 of 436 |
| Time in service | 14 months |
| Agentic actions enabled anywhere on the account | 0 |
| Assistants retired or rebuilt | 0 |
Figures from CustomGPT.ai product usage records. Shared anonymously at the customer’s request.